fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 #undef SCRAMBLE_DELAYED_REFS
  38
  39 /*
  40  * control flags for do_chunk_alloc's force field
  41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  42  * if we really need one.
  43  *
  44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  45  * if we have very few chunks already allocated.  This is
  46  * used as part of the clustering code to help make sure
  47  * we have a good pool of storage to cluster in, without
  48  * filling the FS with empty chunks
  49  *
  50  * CHUNK_ALLOC_FORCE means it must try to allocate one
  51  *
  52  */
  53 enum {
  54         CHUNK_ALLOC_NO_FORCE = 0,
  55         CHUNK_ALLOC_LIMITED = 1,
  56         CHUNK_ALLOC_FORCE = 2,
  57 };
  58
  59 /*
  60  * Control how reservations are dealt with.
  61  *
  62  * RESERVE_FREE - freeing a reservation.
  63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  64  *   ENOSPC accounting
  65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  67  */
  68 enum {
  69         RESERVE_FREE = 0,
  70         RESERVE_ALLOC = 1,
  71         RESERVE_ALLOC_NO_ACCOUNT = 2,
  72 };
  73
  74 static int update_block_group(struct btrfs_trans_handle *trans,
  75                               struct btrfs_root *root,
  76                               u64 bytenr, u64 num_bytes, int alloc);
  77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  78                                 struct btrfs_root *root,
  79                                 u64 bytenr, u64 num_bytes, u64 parent,
  80                                 u64 root_objectid, u64 owner_objectid,
  81                                 u64 owner_offset, int refs_to_drop,
  82                                 struct btrfs_delayed_extent_op *extra_op);
  83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  84                                     struct extent_buffer *leaf,
  85                                     struct btrfs_extent_item *ei);
  86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  87                                       struct btrfs_root *root,
  88                                       u64 parent, u64 root_objectid,
  89                                       u64 flags, u64 owner, u64 offset,
  90                                       struct btrfs_key *ins, int ref_mod);
  91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  92                                      struct btrfs_root *root,
  93                                      u64 parent, u64 root_objectid,
  94                                      u64 flags, struct btrfs_disk_key *key,
  95                                      int level, struct btrfs_key *ins);
  96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  97                           struct btrfs_root *extent_root, u64 alloc_bytes,
  98                           u64 flags, int force);
  99 static int find_next_key(struct btrfs_path *path, int level,
 100                          struct btrfs_key *key);
 101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 102                             int dump_block_groups);
 103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 104                                        u64 num_bytes, int reserve);
 105
 106 static noinline int
 107 block_group_cache_done(struct btrfs_block_group_cache *cache)
 108 {
 109         smp_mb();
 110         return cache->cached == BTRFS_CACHE_FINISHED;
 111 }
 112
 113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 114 {
 115         return (cache->flags & bits) == bits;
 116 }
 117
 118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 119 {
 120         atomic_inc(&cache->count);
 121 }
 122
 123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         if (atomic_dec_and_test(&cache->count)) {
 126                 WARN_ON(cache->pinned > 0);
 127                 WARN_ON(cache->reserved > 0);
 128                 kfree(cache->free_space_ctl);
 129                 kfree(cache);
 130         }
 131 }
 132
 133 /*
 134  * this adds the block group to the fs_info rb tree for the block group
 135  * cache
 136  */
 137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 138                                 struct btrfs_block_group_cache *block_group)
 139 {
 140         struct rb_node **p;
 141         struct rb_node *parent = NULL;
 142         struct btrfs_block_group_cache *cache;
 143
 144         spin_lock(&info->block_group_cache_lock);
 145         p = &info->block_group_cache_tree.rb_node;
 146
 147         while (*p) {
 148                 parent = *p;
 149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 150                                  cache_node);
 151                 if (block_group->key.objectid < cache->key.objectid) {
 152                         p = &(*p)->rb_left;
 153                 } else if (block_group->key.objectid > cache->key.objectid) {
 154                         p = &(*p)->rb_right;
 155                 } else {
 156                         spin_unlock(&info->block_group_cache_lock);
 157                         return -EEXIST;
 158                 }
 159         }
 160
 161         rb_link_node(&block_group->cache_node, parent, p);
 162         rb_insert_color(&block_group->cache_node,
 163                         &info->block_group_cache_tree);
 164         spin_unlock(&info->block_group_cache_lock);
 165
 166         return 0;
 167 }
 168
 169 /*
 170  * This will return the block group at or after bytenr if contains is 0, else
 171  * it will return the block group that contains the bytenr
 172  */
 173 static struct btrfs_block_group_cache *
 174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 175                               int contains)
 176 {
 177         struct btrfs_block_group_cache *cache, *ret = NULL;
 178         struct rb_node *n;
 179         u64 end, start;
 180
 181         spin_lock(&info->block_group_cache_lock);
 182         n = info->block_group_cache_tree.rb_node;
 183
 184         while (n) {
 185                 cache = rb_entry(n, struct btrfs_block_group_cache,
 186                                  cache_node);
 187                 end = cache->key.objectid + cache->key.offset - 1;
 188                 start = cache->key.objectid;
 189
 190                 if (bytenr < start) {
 191                         if (!contains && (!ret || start < ret->key.objectid))
 192                                 ret = cache;
 193                         n = n->rb_left;
 194                 } else if (bytenr > start) {
 195                         if (contains && bytenr <= end) {
 196                                 ret = cache;
 197                                 break;
 198                         }
 199                         n = n->rb_right;
 200                 } else {
 201                         ret = cache;
 202                         break;
 203                 }
 204         }
 205         if (ret)
 206                 btrfs_get_block_group(ret);
 207         spin_unlock(&info->block_group_cache_lock);
 208
 209         return ret;
 210 }
 211
 212 static int add_excluded_extent(struct btrfs_root *root,
 213                                u64 start, u64 num_bytes)
 214 {
 215         u64 end = start + num_bytes - 1;
 216         set_extent_bits(&root->fs_info->freed_extents[0],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         set_extent_bits(&root->fs_info->freed_extents[1],
 219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 220         return 0;
 221 }
 222
 223 static void free_excluded_extents(struct btrfs_root *root,
 224                                   struct btrfs_block_group_cache *cache)
 225 {
 226         u64 start, end;
 227
 228         start = cache->key.objectid;
 229         end = start + cache->key.offset - 1;
 230
 231         clear_extent_bits(&root->fs_info->freed_extents[0],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233         clear_extent_bits(&root->fs_info->freed_extents[1],
 234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 235 }
 236
 237 static int exclude_super_stripes(struct btrfs_root *root,
 238                                  struct btrfs_block_group_cache *cache)
 239 {
 240         u64 bytenr;
 241         u64 *logical;
 242         int stripe_len;
 243         int i, nr, ret;
 244
 245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 247                 cache->bytes_super += stripe_len;
 248                 ret = add_excluded_extent(root, cache->key.objectid,
 249                                           stripe_len);
 250                 BUG_ON(ret); /* -ENOMEM */
 251         }
 252
 253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 254                 bytenr = btrfs_sb_offset(i);
 255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 256                                        cache->key.objectid, bytenr,
 257                                        0, &logical, &nr, &stripe_len);
 258                 BUG_ON(ret); /* -ENOMEM */
 259
 260                 while (nr--) {
 261                         cache->bytes_super += stripe_len;
 262                         ret = add_excluded_extent(root, logical[nr],
 263                                                   stripe_len);
 264                         BUG_ON(ret); /* -ENOMEM */
 265                 }
 266
 267                 kfree(logical);
 268         }
 269         return 0;
 270 }
 271
 272 static struct btrfs_caching_control *
 273 get_caching_control(struct btrfs_block_group_cache *cache)
 274 {
 275         struct btrfs_caching_control *ctl;
 276
 277         spin_lock(&cache->lock);
 278         if (cache->cached != BTRFS_CACHE_STARTED) {
 279                 spin_unlock(&cache->lock);
 280                 return NULL;
 281         }
 282
 283         /* We're loading it the fast way, so we don't have a caching_ctl. */
 284         if (!cache->caching_ctl) {
 285                 spin_unlock(&cache->lock);
 286                 return NULL;
 287         }
 288
 289         ctl = cache->caching_ctl;
 290         atomic_inc(&ctl->count);
 291         spin_unlock(&cache->lock);
 292         return ctl;
 293 }
 294
 295 static void put_caching_control(struct btrfs_caching_control *ctl)
 296 {
 297         if (atomic_dec_and_test(&ctl->count))
 298                 kfree(ctl);
 299 }
 300
 301 /*
 302  * this is only called by cache_block_group, since we could have freed extents
 303  * we need to check the pinned_extents for any extents that can't be used yet
 304  * since their free space will be released as soon as the transaction commits.
 305  */
 306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 307                               struct btrfs_fs_info *info, u64 start, u64 end)
 308 {
 309         u64 extent_start, extent_end, size, total_added = 0;
 310         int ret;
 311
 312         while (start < end) {
 313                 ret = find_first_extent_bit(info->pinned_extents, start,
 314                                             &extent_start, &extent_end,
 315                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 316                 if (ret)
 317                         break;
 318
 319                 if (extent_start <= start) {
 320                         start = extent_end + 1;
 321                 } else if (extent_start > start && extent_start < end) {
 322                         size = extent_start - start;
 323                         total_added += size;
 324                         ret = btrfs_add_free_space(block_group, start,
 325                                                    size);
 326                         BUG_ON(ret); /* -ENOMEM or logic error */
 327                         start = extent_end + 1;
 328                 } else {
 329                         break;
 330                 }
 331         }
 332
 333         if (start < end) {
 334                 size = end - start;
 335                 total_added += size;
 336                 ret = btrfs_add_free_space(block_group, start, size);
 337                 BUG_ON(ret); /* -ENOMEM or logic error */
 338         }
 339
 340         return total_added;
 341 }
 342
 343 static noinline void caching_thread(struct btrfs_work *work)
 344 {
 345         struct btrfs_block_group_cache *block_group;
 346         struct btrfs_fs_info *fs_info;
 347         struct btrfs_caching_control *caching_ctl;
 348         struct btrfs_root *extent_root;
 349         struct btrfs_path *path;
 350         struct extent_buffer *leaf;
 351         struct btrfs_key key;
 352         u64 total_found = 0;
 353         u64 last = 0;
 354         u32 nritems;
 355         int ret = 0;
 356
 357         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 358         block_group = caching_ctl->block_group;
 359         fs_info = block_group->fs_info;
 360         extent_root = fs_info->extent_root;
 361
 362         path = btrfs_alloc_path();
 363         if (!path)
 364                 goto out;
 365
 366         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 367
 368         /*
 369          * We don't want to deadlock with somebody trying to allocate a new
 370          * extent for the extent root while also trying to search the extent
 371          * root to add free space.  So we skip locking and search the commit
 372          * root, since its read-only
 373          */
 374         path->skip_locking = 1;
 375         path->search_commit_root = 1;
 376         path->reada = 1;
 377
 378         key.objectid = last;
 379         key.offset = 0;
 380         key.type = BTRFS_EXTENT_ITEM_KEY;
 381 again:
 382         mutex_lock(&caching_ctl->mutex);
 383         /* need to make sure the commit_root doesn't disappear */
 384         down_read(&fs_info->extent_commit_sem);
 385
 386         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 387         if (ret < 0)
 388                 goto err;
 389
 390         leaf = path->nodes[0];
 391         nritems = btrfs_header_nritems(leaf);
 392
 393         while (1) {
 394                 if (btrfs_fs_closing(fs_info) > 1) {
 395                         last = (u64)-1;
 396                         break;
 397                 }
 398
 399                 if (path->slots[0] < nritems) {
 400                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 401                 } else {
 402                         ret = find_next_key(path, 0, &key);
 403                         if (ret)
 404                                 break;
 405
 406                         if (need_resched() ||
 407                             btrfs_next_leaf(extent_root, path)) {
 408                                 caching_ctl->progress = last;
 409                                 btrfs_release_path(path);
 410                                 up_read(&fs_info->extent_commit_sem);
 411                                 mutex_unlock(&caching_ctl->mutex);
 412                                 cond_resched();
 413                                 goto again;
 414                         }
 415                         leaf = path->nodes[0];
 416                         nritems = btrfs_header_nritems(leaf);
 417                         continue;
 418                 }
 419
 420                 if (key.objectid < block_group->key.objectid) {
 421                         path->slots[0]++;
 422                         continue;
 423                 }
 424
 425                 if (key.objectid >= block_group->key.objectid +
 426                     block_group->key.offset)
 427                         break;
 428
 429                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 430                         total_found += add_new_free_space(block_group,
 431                                                           fs_info, last,
 432                                                           key.objectid);
 433                         last = key.objectid + key.offset;
 434
 435                         if (total_found > (1024 * 1024 * 2)) {
 436                                 total_found = 0;
 437                                 wake_up(&caching_ctl->wait);
 438                         }
 439                 }
 440                 path->slots[0]++;
 441         }
 442         ret = 0;
 443
 444         total_found += add_new_free_space(block_group, fs_info, last,
 445                                           block_group->key.objectid +
 446                                           block_group->key.offset);
 447         caching_ctl->progress = (u64)-1;
 448
 449         spin_lock(&block_group->lock);
 450         block_group->caching_ctl = NULL;
 451         block_group->cached = BTRFS_CACHE_FINISHED;
 452         spin_unlock(&block_group->lock);
 453
 454 err:
 455         btrfs_free_path(path);
 456         up_read(&fs_info->extent_commit_sem);
 457
 458         free_excluded_extents(extent_root, block_group);
 459
 460         mutex_unlock(&caching_ctl->mutex);
 461 out:
 462         wake_up(&caching_ctl->wait);
 463
 464         put_caching_control(caching_ctl);
 465         btrfs_put_block_group(block_group);
 466 }
 467
 468 static int cache_block_group(struct btrfs_block_group_cache *cache,
 469                              struct btrfs_trans_handle *trans,
 470                              struct btrfs_root *root,
 471                              int load_cache_only)
 472 {
 473         DEFINE_WAIT(wait);
 474         struct btrfs_fs_info *fs_info = cache->fs_info;
 475         struct btrfs_caching_control *caching_ctl;
 476         int ret = 0;
 477
 478         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 479         if (!caching_ctl)
 480                 return -ENOMEM;
 481
 482         INIT_LIST_HEAD(&caching_ctl->list);
 483         mutex_init(&caching_ctl->mutex);
 484         init_waitqueue_head(&caching_ctl->wait);
 485         caching_ctl->block_group = cache;
 486         caching_ctl->progress = cache->key.objectid;
 487         atomic_set(&caching_ctl->count, 1);
 488         caching_ctl->work.func = caching_thread;
 489
 490         spin_lock(&cache->lock);
 491         /*
 492          * This should be a rare occasion, but this could happen I think in the
 493          * case where one thread starts to load the space cache info, and then
 494          * some other thread starts a transaction commit which tries to do an
 495          * allocation while the other thread is still loading the space cache
 496          * info.  The previous loop should have kept us from choosing this block
 497          * group, but if we've moved to the state where we will wait on caching
 498          * block groups we need to first check if we're doing a fast load here,
 499          * so we can wait for it to finish, otherwise we could end up allocating
 500          * from a block group who's cache gets evicted for one reason or
 501          * another.
 502          */
 503         while (cache->cached == BTRFS_CACHE_FAST) {
 504                 struct btrfs_caching_control *ctl;
 505
 506                 ctl = cache->caching_ctl;
 507                 atomic_inc(&ctl->count);
 508                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 509                 spin_unlock(&cache->lock);
 510
 511                 schedule();
 512
 513                 finish_wait(&ctl->wait, &wait);
 514                 put_caching_control(ctl);
 515                 spin_lock(&cache->lock);
 516         }
 517
 518         if (cache->cached != BTRFS_CACHE_NO) {
 519                 spin_unlock(&cache->lock);
 520                 kfree(caching_ctl);
 521                 return 0;
 522         }
 523         WARN_ON(cache->caching_ctl);
 524         cache->caching_ctl = caching_ctl;
 525         cache->cached = BTRFS_CACHE_FAST;
 526         spin_unlock(&cache->lock);
 527
 528         /*
 529          * We can't do the read from on-disk cache during a commit since we need
 530          * to have the normal tree locking.  Also if we are currently trying to
 531          * allocate blocks for the tree root we can't do the fast caching since
 532          * we likely hold important locks.
 533          */
 534         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 535                 ret = load_free_space_cache(fs_info, cache);
 536
 537                 spin_lock(&cache->lock);
 538                 if (ret == 1) {
 539                         cache->caching_ctl = NULL;
 540                         cache->cached = BTRFS_CACHE_FINISHED;
 541                         cache->last_byte_to_unpin = (u64)-1;
 542                 } else {
 543                         if (load_cache_only) {
 544                                 cache->caching_ctl = NULL;
 545                                 cache->cached = BTRFS_CACHE_NO;
 546                         } else {
 547                                 cache->cached = BTRFS_CACHE_STARTED;
 548                         }
 549                 }
 550                 spin_unlock(&cache->lock);
 551                 wake_up(&caching_ctl->wait);
 552                 if (ret == 1) {
 553                         put_caching_control(caching_ctl);
 554                         free_excluded_extents(fs_info->extent_root, cache);
 555                         return 0;
 556                 }
 557         } else {
 558                 /*
 559                  * We are not going to do the fast caching, set cached to the
 560                  * appropriate value and wakeup any waiters.
 561                  */
 562                 spin_lock(&cache->lock);
 563                 if (load_cache_only) {
 564                         cache->caching_ctl = NULL;
 565                         cache->cached = BTRFS_CACHE_NO;
 566                 } else {
 567                         cache->cached = BTRFS_CACHE_STARTED;
 568                 }
 569                 spin_unlock(&cache->lock);
 570                 wake_up(&caching_ctl->wait);
 571         }
 572
 573         if (load_cache_only) {
 574                 put_caching_control(caching_ctl);
 575                 return 0;
 576         }
 577
 578         down_write(&fs_info->extent_commit_sem);
 579         atomic_inc(&caching_ctl->count);
 580         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 581         up_write(&fs_info->extent_commit_sem);
 582
 583         btrfs_get_block_group(cache);
 584
 585         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 586
 587         return ret;
 588 }
 589
 590 /*
 591  * return the block group that starts at or after bytenr
 592  */
 593 static struct btrfs_block_group_cache *
 594 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 595 {
 596         struct btrfs_block_group_cache *cache;
 597
 598         cache = block_group_cache_tree_search(info, bytenr, 0);
 599
 600         return cache;
 601 }
 602
 603 /*
 604  * return the block group that contains the given bytenr
 605  */
 606 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 607                                                  struct btrfs_fs_info *info,
 608                                                  u64 bytenr)
 609 {
 610         struct btrfs_block_group_cache *cache;
 611
 612         cache = block_group_cache_tree_search(info, bytenr, 1);
 613
 614         return cache;
 615 }
 616
 617 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 618                                                   u64 flags)
 619 {
 620         struct list_head *head = &info->space_info;
 621         struct btrfs_space_info *found;
 622
 623         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 624
 625         rcu_read_lock();
 626         list_for_each_entry_rcu(found, head, list) {
 627                 if (found->flags & flags) {
 628                         rcu_read_unlock();
 629                         return found;
 630                 }
 631         }
 632         rcu_read_unlock();
 633         return NULL;
 634 }
 635
 636 /*
 637  * after adding space to the filesystem, we need to clear the full flags
 638  * on all the space infos.
 639  */
 640 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 641 {
 642         struct list_head *head = &info->space_info;
 643         struct btrfs_space_info *found;
 644
 645         rcu_read_lock();
 646         list_for_each_entry_rcu(found, head, list)
 647                 found->full = 0;
 648         rcu_read_unlock();
 649 }
 650
 651 static u64 div_factor(u64 num, int factor)
 652 {
 653         if (factor == 10)
 654                 return num;
 655         num *= factor;
 656         do_div(num, 10);
 657         return num;
 658 }
 659
 660 static u64 div_factor_fine(u64 num, int factor)
 661 {
 662         if (factor == 100)
 663                 return num;
 664         num *= factor;
 665         do_div(num, 100);
 666         return num;
 667 }
 668
 669 u64 btrfs_find_block_group(struct btrfs_root *root,
 670                            u64 search_start, u64 search_hint, int owner)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673         u64 used;
 674         u64 last = max(search_hint, search_start);
 675         u64 group_start = 0;
 676         int full_search = 0;
 677         int factor = 9;
 678         int wrapped = 0;
 679 again:
 680         while (1) {
 681                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 682                 if (!cache)
 683                         break;
 684
 685                 spin_lock(&cache->lock);
 686                 last = cache->key.objectid + cache->key.offset;
 687                 used = btrfs_block_group_used(&cache->item);
 688
 689                 if ((full_search || !cache->ro) &&
 690                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 691                         if (used + cache->pinned + cache->reserved <
 692                             div_factor(cache->key.offset, factor)) {
 693                                 group_start = cache->key.objectid;
 694                                 spin_unlock(&cache->lock);
 695                                 btrfs_put_block_group(cache);
 696                                 goto found;
 697                         }
 698                 }
 699                 spin_unlock(&cache->lock);
 700                 btrfs_put_block_group(cache);
 701                 cond_resched();
 702         }
 703         if (!wrapped) {
 704                 last = search_start;
 705                 wrapped = 1;
 706                 goto again;
 707         }
 708         if (!full_search && factor < 10) {
 709                 last = search_start;
 710                 full_search = 1;
 711                 factor = 10;
 712                 goto again;
 713         }
 714 found:
 715         return group_start;
 716 }
 717
 718 /* simple helper to search for an existing extent at a given offset */
 719 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 720 {
 721         int ret;
 722         struct btrfs_key key;
 723         struct btrfs_path *path;
 724
 725         path = btrfs_alloc_path();
 726         if (!path)
 727                 return -ENOMEM;
 728
 729         key.objectid = start;
 730         key.offset = len;
 731         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 732         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 733                                 0, 0);
 734         btrfs_free_path(path);
 735         return ret;
 736 }
 737
 738 /*
 739  * helper function to lookup reference count and flags of extent.
 740  *
 741  * the head node for delayed ref is used to store the sum of all the
 742  * reference count modifications queued up in the rbtree. the head
 743  * node may also store the extent flags to set. This way you can check
 744  * to see what the reference count and extent flags would be if all of
 745  * the delayed refs are not processed.
 746  */
 747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 748                              struct btrfs_root *root, u64 bytenr,
 749                              u64 num_bytes, u64 *refs, u64 *flags)
 750 {
 751         struct btrfs_delayed_ref_head *head;
 752         struct btrfs_delayed_ref_root *delayed_refs;
 753         struct btrfs_path *path;
 754         struct btrfs_extent_item *ei;
 755         struct extent_buffer *leaf;
 756         struct btrfs_key key;
 757         u32 item_size;
 758         u64 num_refs;
 759         u64 extent_flags;
 760         int ret;
 761
 762         path = btrfs_alloc_path();
 763         if (!path)
 764                 return -ENOMEM;
 765
 766         key.objectid = bytenr;
 767         key.type = BTRFS_EXTENT_ITEM_KEY;
 768         key.offset = num_bytes;
 769         if (!trans) {
 770                 path->skip_locking = 1;
 771                 path->search_commit_root = 1;
 772         }
 773 again:
 774         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 775                                 &key, path, 0, 0);
 776         if (ret < 0)
 777                 goto out_free;
 778
 779         if (ret == 0) {
 780                 leaf = path->nodes[0];
 781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 782                 if (item_size >= sizeof(*ei)) {
 783                         ei = btrfs_item_ptr(leaf, path->slots[0],
 784                                             struct btrfs_extent_item);
 785                         num_refs = btrfs_extent_refs(leaf, ei);
 786                         extent_flags = btrfs_extent_flags(leaf, ei);
 787                 } else {
 788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 789                         struct btrfs_extent_item_v0 *ei0;
 790                         BUG_ON(item_size != sizeof(*ei0));
 791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 792                                              struct btrfs_extent_item_v0);
 793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 794                         /* FIXME: this isn't correct for data */
 795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 796 #else
 797                         BUG();
 798 #endif
 799                 }
 800                 BUG_ON(num_refs == 0);
 801         } else {
 802                 num_refs = 0;
 803                 extent_flags = 0;
 804                 ret = 0;
 805         }
 806
 807         if (!trans)
 808                 goto out;
 809
 810         delayed_refs = &trans->transaction->delayed_refs;
 811         spin_lock(&delayed_refs->lock);
 812         head = btrfs_find_delayed_ref_head(trans, bytenr);
 813         if (head) {
 814                 if (!mutex_trylock(&head->mutex)) {
 815                         atomic_inc(&head->node.refs);
 816                         spin_unlock(&delayed_refs->lock);
 817
 818                         btrfs_release_path(path);
 819
 820                         /*
 821                          * Mutex was contended, block until it's released and try
 822                          * again
 823                          */
 824                         mutex_lock(&head->mutex);
 825                         mutex_unlock(&head->mutex);
 826                         btrfs_put_delayed_ref(&head->node);
 827                         goto again;
 828                 }
 829                 if (head->extent_op && head->extent_op->update_flags)
 830                         extent_flags |= head->extent_op->flags_to_set;
 831                 else
 832                         BUG_ON(num_refs == 0);
 833
 834                 num_refs += head->node.ref_mod;
 835                 mutex_unlock(&head->mutex);
 836         }
 837         spin_unlock(&delayed_refs->lock);
 838 out:
 839         WARN_ON(num_refs == 0);
 840         if (refs)
 841                 *refs = num_refs;
 842         if (flags)
 843                 *flags = extent_flags;
 844 out_free:
 845         btrfs_free_path(path);
 846         return ret;
 847 }
 848
 849 /*
 850  * Back reference rules.  Back refs have three main goals:
 851  *
 852  * 1) differentiate between all holders of references to an extent so that
 853  *    when a reference is dropped we can make sure it was a valid reference
 854  *    before freeing the extent.
 855  *
 856  * 2) Provide enough information to quickly find the holders of an extent
 857  *    if we notice a given block is corrupted or bad.
 858  *
 859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 860  *    maintenance.  This is actually the same as #2, but with a slightly
 861  *    different use case.
 862  *
 863  * There are two kinds of back refs. The implicit back refs is optimized
 864  * for pointers in non-shared tree blocks. For a given pointer in a block,
 865  * back refs of this kind provide information about the block's owner tree
 866  * and the pointer's key. These information allow us to find the block by
 867  * b-tree searching. The full back refs is for pointers in tree blocks not
 868  * referenced by their owner trees. The location of tree block is recorded
 869  * in the back refs. Actually the full back refs is generic, and can be
 870  * used in all cases the implicit back refs is used. The major shortcoming
 871  * of the full back refs is its overhead. Every time a tree block gets
 872  * COWed, we have to update back refs entry for all pointers in it.
 873  *
 874  * For a newly allocated tree block, we use implicit back refs for
 875  * pointers in it. This means most tree related operations only involve
 876  * implicit back refs. For a tree block created in old transaction, the
 877  * only way to drop a reference to it is COW it. So we can detect the
 878  * event that tree block loses its owner tree's reference and do the
 879  * back refs conversion.
 880  *
 881  * When a tree block is COW'd through a tree, there are four cases:
 882  *
 883  * The reference count of the block is one and the tree is the block's
 884  * owner tree. Nothing to do in this case.
 885  *
 886  * The reference count of the block is one and the tree is not the
 887  * block's owner tree. In this case, full back refs is used for pointers
 888  * in the block. Remove these full back refs, add implicit back refs for
 889  * every pointers in the new block.
 890  *
 891  * The reference count of the block is greater than one and the tree is
 892  * the block's owner tree. In this case, implicit back refs is used for
 893  * pointers in the block. Add full back refs for every pointers in the
 894  * block, increase lower level extents' reference counts. The original
 895  * implicit back refs are entailed to the new block.
 896  *
 897  * The reference count of the block is greater than one and the tree is
 898  * not the block's owner tree. Add implicit back refs for every pointer in
 899  * the new block, increase lower level extents' reference count.
 900  *
 901  * Back Reference Key composing:
 902  *
 903  * The key objectid corresponds to the first byte in the extent,
 904  * The key type is used to differentiate between types of back refs.
 905  * There are different meanings of the key offset for different types
 906  * of back refs.
 907  *
 908  * File extents can be referenced by:
 909  *
 910  * - multiple snapshots, subvolumes, or different generations in one subvol
 911  * - different files inside a single subvolume
 912  * - different offsets inside a file (bookend extents in file.c)
 913  *
 914  * The extent ref structure for the implicit back refs has fields for:
 915  *
 916  * - Objectid of the subvolume root
 917  * - objectid of the file holding the reference
 918  * - original offset in the file
 919  * - how many bookend extents
 920  *
 921  * The key offset for the implicit back refs is hash of the first
 922  * three fields.
 923  *
 924  * The extent ref structure for the full back refs has field for:
 925  *
 926  * - number of pointers in the tree leaf
 927  *
 928  * The key offset for the implicit back refs is the first byte of
 929  * the tree leaf
 930  *
 931  * When a file extent is allocated, The implicit back refs is used.
 932  * the fields are filled in:
 933  *
 934  *     (root_key.objectid, inode objectid, offset in file, 1)
 935  *
 936  * When a file extent is removed file truncation, we find the
 937  * corresponding implicit back refs and check the following fields:
 938  *
 939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 940  *
 941  * Btree extents can be referenced by:
 942  *
 943  * - Different subvolumes
 944  *
 945  * Both the implicit back refs and the full back refs for tree blocks
 946  * only consist of key. The key offset for the implicit back refs is
 947  * objectid of block's owner tree. The key offset for the full back refs
 948  * is the first byte of parent block.
 949  *
 950  * When implicit back refs is used, information about the lowest key and
 951  * level of the tree block are required. These information are stored in
 952  * tree block info structure.
 953  */
 954
 955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 957                                   struct btrfs_root *root,
 958                                   struct btrfs_path *path,
 959                                   u64 owner, u32 extra_size)
 960 {
 961         struct btrfs_extent_item *item;
 962         struct btrfs_extent_item_v0 *ei0;
 963         struct btrfs_extent_ref_v0 *ref0;
 964         struct btrfs_tree_block_info *bi;
 965         struct extent_buffer *leaf;
 966         struct btrfs_key key;
 967         struct btrfs_key found_key;
 968         u32 new_size = sizeof(*item);
 969         u64 refs;
 970         int ret;
 971
 972         leaf = path->nodes[0];
 973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 974
 975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 977                              struct btrfs_extent_item_v0);
 978         refs = btrfs_extent_refs_v0(leaf, ei0);
 979
 980         if (owner == (u64)-1) {
 981                 while (1) {
 982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 983                                 ret = btrfs_next_leaf(root, path);
 984                                 if (ret < 0)
 985                                         return ret;
 986                                 BUG_ON(ret > 0); /* Corruption */
 987                                 leaf = path->nodes[0];
 988                         }
 989                         btrfs_item_key_to_cpu(leaf, &found_key,
 990                                               path->slots[0]);
 991                         BUG_ON(key.objectid != found_key.objectid);
 992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 993                                 path->slots[0]++;
 994                                 continue;
 995                         }
 996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 997                                               struct btrfs_extent_ref_v0);
 998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(trans, root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466
1467         key.objectid = bytenr;
1468         key.type = BTRFS_EXTENT_ITEM_KEY;
1469         key.offset = num_bytes;
1470
1471         want = extent_ref_type(parent, owner);
1472         if (insert) {
1473                 extra_size = btrfs_extent_inline_ref_size(want);
1474                 path->keep_locks = 1;
1475         } else
1476                 extra_size = -1;
1477         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1478         if (ret < 0) {
1479                 err = ret;
1480                 goto out;
1481         }
1482         if (ret && !insert) {
1483                 err = -ENOENT;
1484                 goto out;
1485         }
1486         BUG_ON(ret); /* Corruption */
1487
1488         leaf = path->nodes[0];
1489         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1490 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1491         if (item_size < sizeof(*ei)) {
1492                 if (!insert) {
1493                         err = -ENOENT;
1494                         goto out;
1495                 }
1496                 ret = convert_extent_item_v0(trans, root, path, owner,
1497                                              extra_size);
1498                 if (ret < 0) {
1499                         err = ret;
1500                         goto out;
1501                 }
1502                 leaf = path->nodes[0];
1503                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1504         }
1505 #endif
1506         BUG_ON(item_size < sizeof(*ei));
1507
1508         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1509         flags = btrfs_extent_flags(leaf, ei);
1510
1511         ptr = (unsigned long)(ei + 1);
1512         end = (unsigned long)ei + item_size;
1513
1514         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1515                 ptr += sizeof(struct btrfs_tree_block_info);
1516                 BUG_ON(ptr > end);
1517         } else {
1518                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1519         }
1520
1521         err = -ENOENT;
1522         while (1) {
1523                 if (ptr >= end) {
1524                         WARN_ON(ptr > end);
1525                         break;
1526                 }
1527                 iref = (struct btrfs_extent_inline_ref *)ptr;
1528                 type = btrfs_extent_inline_ref_type(leaf, iref);
1529                 if (want < type)
1530                         break;
1531                 if (want > type) {
1532                         ptr += btrfs_extent_inline_ref_size(type);
1533                         continue;
1534                 }
1535
1536                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1537                         struct btrfs_extent_data_ref *dref;
1538                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1539                         if (match_extent_data_ref(leaf, dref, root_objectid,
1540                                                   owner, offset)) {
1541                                 err = 0;
1542                                 break;
1543                         }
1544                         if (hash_extent_data_ref_item(leaf, dref) <
1545                             hash_extent_data_ref(root_objectid, owner, offset))
1546                                 break;
1547                 } else {
1548                         u64 ref_offset;
1549                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1550                         if (parent > 0) {
1551                                 if (parent == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < parent)
1556                                         break;
1557                         } else {
1558                                 if (root_objectid == ref_offset) {
1559                                         err = 0;
1560                                         break;
1561                                 }
1562                                 if (ref_offset < root_objectid)
1563                                         break;
1564                         }
1565                 }
1566                 ptr += btrfs_extent_inline_ref_size(type);
1567         }
1568         if (err == -ENOENT && insert) {
1569                 if (item_size + extra_size >=
1570                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574                 /*
1575                  * To add new inline back ref, we have to make sure
1576                  * there is no corresponding back ref item.
1577                  * For simplicity, we just do not add new inline back
1578                  * ref if there is any kind of item for this block
1579                  */
1580                 if (find_next_key(path, 0, &key) == 0 &&
1581                     key.objectid == bytenr &&
1582                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1583                         err = -EAGAIN;
1584                         goto out;
1585                 }
1586         }
1587         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1588 out:
1589         if (insert) {
1590                 path->keep_locks = 0;
1591                 btrfs_unlock_up_safe(path, 1);
1592         }
1593         return err;
1594 }
1595
1596 /*
1597  * helper to add new inline back ref
1598  */
1599 static noinline_for_stack
1600 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_root *root,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref *iref,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int refs_to_add,
1606                                  struct btrfs_delayed_extent_op *extent_op)
1607 {
1608         struct extent_buffer *leaf;
1609         struct btrfs_extent_item *ei;
1610         unsigned long ptr;
1611         unsigned long end;
1612         unsigned long item_offset;
1613         u64 refs;
1614         int size;
1615         int type;
1616
1617         leaf = path->nodes[0];
1618         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1619         item_offset = (unsigned long)iref - (unsigned long)ei;
1620
1621         type = extent_ref_type(parent, owner);
1622         size = btrfs_extent_inline_ref_size(type);
1623
1624         btrfs_extend_item(trans, root, path, size);
1625
1626         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1627         refs = btrfs_extent_refs(leaf, ei);
1628         refs += refs_to_add;
1629         btrfs_set_extent_refs(leaf, ei, refs);
1630         if (extent_op)
1631                 __run_delayed_extent_op(extent_op, leaf, ei);
1632
1633         ptr = (unsigned long)ei + item_offset;
1634         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1635         if (ptr < end - size)
1636                 memmove_extent_buffer(leaf, ptr + size, ptr,
1637                                       end - size - ptr);
1638
1639         iref = (struct btrfs_extent_inline_ref *)ptr;
1640         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1641         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1642                 struct btrfs_extent_data_ref *dref;
1643                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1644                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1645                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1646                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1647                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1648         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1649                 struct btrfs_shared_data_ref *sref;
1650                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1651                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1655         } else {
1656                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1657         }
1658         btrfs_mark_buffer_dirty(leaf);
1659 }
1660
1661 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1662                                  struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref **ref_ret,
1665                                  u64 bytenr, u64 num_bytes, u64 parent,
1666                                  u64 root_objectid, u64 owner, u64 offset)
1667 {
1668         int ret;
1669
1670         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1671                                            bytenr, num_bytes, parent,
1672                                            root_objectid, owner, offset, 0);
1673         if (ret != -ENOENT)
1674                 return ret;
1675
1676         btrfs_release_path(path);
1677         *ref_ret = NULL;
1678
1679         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1680                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1681                                             root_objectid);
1682         } else {
1683                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1684                                              root_objectid, owner, offset);
1685         }
1686         return ret;
1687 }
1688
1689 /*
1690  * helper to update/remove inline back ref
1691  */
1692 static noinline_for_stack
1693 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1694                                   struct btrfs_root *root,
1695                                   struct btrfs_path *path,
1696                                   struct btrfs_extent_inline_ref *iref,
1697                                   int refs_to_mod,
1698                                   struct btrfs_delayed_extent_op *extent_op)
1699 {
1700         struct extent_buffer *leaf;
1701         struct btrfs_extent_item *ei;
1702         struct btrfs_extent_data_ref *dref = NULL;
1703         struct btrfs_shared_data_ref *sref = NULL;
1704         unsigned long ptr;
1705         unsigned long end;
1706         u32 item_size;
1707         int size;
1708         int type;
1709         u64 refs;
1710
1711         leaf = path->nodes[0];
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1715         refs += refs_to_mod;
1716         btrfs_set_extent_refs(leaf, ei, refs);
1717         if (extent_op)
1718                 __run_delayed_extent_op(extent_op, leaf, ei);
1719
1720         type = btrfs_extent_inline_ref_type(leaf, iref);
1721
1722         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1723                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1724                 refs = btrfs_extent_data_ref_count(leaf, dref);
1725         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1726                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1727                 refs = btrfs_shared_data_ref_count(leaf, sref);
1728         } else {
1729                 refs = 1;
1730                 BUG_ON(refs_to_mod != -1);
1731         }
1732
1733         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1734         refs += refs_to_mod;
1735
1736         if (refs > 0) {
1737                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1738                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1739                 else
1740                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1741         } else {
1742                 size =  btrfs_extent_inline_ref_size(type);
1743                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1744                 ptr = (unsigned long)iref;
1745                 end = (unsigned long)ei + item_size;
1746                 if (ptr + size < end)
1747                         memmove_extent_buffer(leaf, ptr, ptr + size,
1748                                               end - ptr - size);
1749                 item_size -= size;
1750                 btrfs_truncate_item(trans, root, path, item_size, 1);
1751         }
1752         btrfs_mark_buffer_dirty(leaf);
1753 }
1754
1755 static noinline_for_stack
1756 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1757                                  struct btrfs_root *root,
1758                                  struct btrfs_path *path,
1759                                  u64 bytenr, u64 num_bytes, u64 parent,
1760                                  u64 root_objectid, u64 owner,
1761                                  u64 offset, int refs_to_add,
1762                                  struct btrfs_delayed_extent_op *extent_op)
1763 {
1764         struct btrfs_extent_inline_ref *iref;
1765         int ret;
1766
1767         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1768                                            bytenr, num_bytes, parent,
1769                                            root_objectid, owner, offset, 1);
1770         if (ret == 0) {
1771                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1772                 update_inline_extent_backref(trans, root, path, iref,
1773                                              refs_to_add, extent_op);
1774         } else if (ret == -ENOENT) {
1775                 setup_inline_extent_backref(trans, root, path, iref, parent,
1776                                             root_objectid, owner, offset,
1777                                             refs_to_add, extent_op);
1778                 ret = 0;
1779         }
1780         return ret;
1781 }
1782
1783 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1784                                  struct btrfs_root *root,
1785                                  struct btrfs_path *path,
1786                                  u64 bytenr, u64 parent, u64 root_objectid,
1787                                  u64 owner, u64 offset, int refs_to_add)
1788 {
1789         int ret;
1790         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1791                 BUG_ON(refs_to_add != 1);
1792                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1793                                             parent, root_objectid);
1794         } else {
1795                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1796                                              parent, root_objectid,
1797                                              owner, offset, refs_to_add);
1798         }
1799         return ret;
1800 }
1801
1802 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  int refs_to_drop, int is_data)
1807 {
1808         int ret = 0;
1809
1810         BUG_ON(!is_data && refs_to_drop != 1);
1811         if (iref) {
1812                 update_inline_extent_backref(trans, root, path, iref,
1813                                              -refs_to_drop, NULL);
1814         } else if (is_data) {
1815                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1816         } else {
1817                 ret = btrfs_del_item(trans, root, path);
1818         }
1819         return ret;
1820 }
1821
1822 static int btrfs_issue_discard(struct block_device *bdev,
1823                                 u64 start, u64 len)
1824 {
1825         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1826 }
1827
1828 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1829                                 u64 num_bytes, u64 *actual_bytes)
1830 {
1831         int ret;
1832         u64 discarded_bytes = 0;
1833         struct btrfs_bio *bbio = NULL;
1834
1835
1836         /* Tell the block device(s) that the sectors can be discarded */
1837         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1838                               bytenr, &num_bytes, &bbio, 0);
1839         /* Error condition is -ENOMEM */
1840         if (!ret) {
1841                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1842                 int i;
1843
1844
1845                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1846                         if (!stripe->dev->can_discard)
1847                                 continue;
1848
1849                         ret = btrfs_issue_discard(stripe->dev->bdev,
1850                                                   stripe->physical,
1851                                                   stripe->length);
1852                         if (!ret)
1853                                 discarded_bytes += stripe->length;
1854                         else if (ret != -EOPNOTSUPP)
1855                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1856
1857                         /*
1858                          * Just in case we get back EOPNOTSUPP for some reason,
1859                          * just ignore the return value so we don't screw up
1860                          * people calling discard_extent.
1861                          */
1862                         ret = 0;
1863                 }
1864                 kfree(bbio);
1865         }
1866
1867         if (actual_bytes)
1868                 *actual_bytes = discarded_bytes;
1869
1870
1871         return ret;
1872 }
1873
1874 /* Can return -ENOMEM */
1875 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          u64 bytenr, u64 num_bytes, u64 parent,
1878                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1879 {
1880         int ret;
1881         struct btrfs_fs_info *fs_info = root->fs_info;
1882
1883         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1884                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1885
1886         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, (int)owner,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         } else {
1892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1893                                         num_bytes,
1894                                         parent, root_objectid, owner, offset,
1895                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1896         }
1897         return ret;
1898 }
1899
1900 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1901                                   struct btrfs_root *root,
1902                                   u64 bytenr, u64 num_bytes,
1903                                   u64 parent, u64 root_objectid,
1904                                   u64 owner, u64 offset, int refs_to_add,
1905                                   struct btrfs_delayed_extent_op *extent_op)
1906 {
1907         struct btrfs_path *path;
1908         struct extent_buffer *leaf;
1909         struct btrfs_extent_item *item;
1910         u64 refs;
1911         int ret;
1912         int err = 0;
1913
1914         path = btrfs_alloc_path();
1915         if (!path)
1916                 return -ENOMEM;
1917
1918         path->reada = 1;
1919         path->leave_spinning = 1;
1920         /* this will setup the path even if it fails to insert the back ref */
1921         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1922                                            path, bytenr, num_bytes, parent,
1923                                            root_objectid, owner, offset,
1924                                            refs_to_add, extent_op);
1925         if (ret == 0)
1926                 goto out;
1927
1928         if (ret != -EAGAIN) {
1929                 err = ret;
1930                 goto out;
1931         }
1932
1933         leaf = path->nodes[0];
1934         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1935         refs = btrfs_extent_refs(leaf, item);
1936         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1937         if (extent_op)
1938                 __run_delayed_extent_op(extent_op, leaf, item);
1939
1940         btrfs_mark_buffer_dirty(leaf);
1941         btrfs_release_path(path);
1942
1943         path->reada = 1;
1944         path->leave_spinning = 1;
1945
1946         /* now insert the actual backref */
1947         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1948                                     path, bytenr, parent, root_objectid,
1949                                     owner, offset, refs_to_add);
1950         if (ret)
1951                 btrfs_abort_transaction(trans, root, ret);
1952 out:
1953         btrfs_free_path(path);
1954         return err;
1955 }
1956
1957 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1958                                 struct btrfs_root *root,
1959                                 struct btrfs_delayed_ref_node *node,
1960                                 struct btrfs_delayed_extent_op *extent_op,
1961                                 int insert_reserved)
1962 {
1963         int ret = 0;
1964         struct btrfs_delayed_data_ref *ref;
1965         struct btrfs_key ins;
1966         u64 parent = 0;
1967         u64 ref_root = 0;
1968         u64 flags = 0;
1969
1970         ins.objectid = node->bytenr;
1971         ins.offset = node->num_bytes;
1972         ins.type = BTRFS_EXTENT_ITEM_KEY;
1973
1974         ref = btrfs_delayed_node_to_data_ref(node);
1975         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1976                 parent = ref->parent;
1977         else
1978                 ref_root = ref->root;
1979
1980         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1981                 if (extent_op) {
1982                         BUG_ON(extent_op->update_key);
1983                         flags |= extent_op->flags_to_set;
1984                 }
1985                 ret = alloc_reserved_file_extent(trans, root,
1986                                                  parent, ref_root, flags,
1987                                                  ref->objectid, ref->offset,
1988                                                  &ins, node->ref_mod);
1989         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1990                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1991                                              node->num_bytes, parent,
1992                                              ref_root, ref->objectid,
1993                                              ref->offset, node->ref_mod,
1994                                              extent_op);
1995         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1996                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1997                                           node->num_bytes, parent,
1998                                           ref_root, ref->objectid,
1999                                           ref->offset, node->ref_mod,
2000                                           extent_op);
2001         } else {
2002                 BUG();
2003         }
2004         return ret;
2005 }
2006
2007 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2008                                     struct extent_buffer *leaf,
2009                                     struct btrfs_extent_item *ei)
2010 {
2011         u64 flags = btrfs_extent_flags(leaf, ei);
2012         if (extent_op->update_flags) {
2013                 flags |= extent_op->flags_to_set;
2014                 btrfs_set_extent_flags(leaf, ei, flags);
2015         }
2016
2017         if (extent_op->update_key) {
2018                 struct btrfs_tree_block_info *bi;
2019                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2020                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2021                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2022         }
2023 }
2024
2025 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2026                                  struct btrfs_root *root,
2027                                  struct btrfs_delayed_ref_node *node,
2028                                  struct btrfs_delayed_extent_op *extent_op)
2029 {
2030         struct btrfs_key key;
2031         struct btrfs_path *path;
2032         struct btrfs_extent_item *ei;
2033         struct extent_buffer *leaf;
2034         u32 item_size;
2035         int ret;
2036         int err = 0;
2037
2038         if (trans->aborted)
2039                 return 0;
2040
2041         path = btrfs_alloc_path();
2042         if (!path)
2043                 return -ENOMEM;
2044
2045         key.objectid = node->bytenr;
2046         key.type = BTRFS_EXTENT_ITEM_KEY;
2047         key.offset = node->num_bytes;
2048
2049         path->reada = 1;
2050         path->leave_spinning = 1;
2051         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2052                                 path, 0, 1);
2053         if (ret < 0) {
2054                 err = ret;
2055                 goto out;
2056         }
2057         if (ret > 0) {
2058                 err = -EIO;
2059                 goto out;
2060         }
2061
2062         leaf = path->nodes[0];
2063         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2065         if (item_size < sizeof(*ei)) {
2066                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2067                                              path, (u64)-1, 0);
2068                 if (ret < 0) {
2069                         err = ret;
2070                         goto out;
2071                 }
2072                 leaf = path->nodes[0];
2073                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2074         }
2075 #endif
2076         BUG_ON(item_size < sizeof(*ei));
2077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         __run_delayed_extent_op(extent_op, leaf, ei);
2079
2080         btrfs_mark_buffer_dirty(leaf);
2081 out:
2082         btrfs_free_path(path);
2083         return err;
2084 }
2085
2086 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_root *root,
2088                                 struct btrfs_delayed_ref_node *node,
2089                                 struct btrfs_delayed_extent_op *extent_op,
2090                                 int insert_reserved)
2091 {
2092         int ret = 0;
2093         struct btrfs_delayed_tree_ref *ref;
2094         struct btrfs_key ins;
2095         u64 parent = 0;
2096         u64 ref_root = 0;
2097
2098         ins.objectid = node->bytenr;
2099         ins.offset = node->num_bytes;
2100         ins.type = BTRFS_EXTENT_ITEM_KEY;
2101
2102         ref = btrfs_delayed_node_to_tree_ref(node);
2103         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2104                 parent = ref->parent;
2105         else
2106                 ref_root = ref->root;
2107
2108         BUG_ON(node->ref_mod != 1);
2109         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2110                 BUG_ON(!extent_op || !extent_op->update_flags ||
2111                        !extent_op->update_key);
2112                 ret = alloc_reserved_tree_block(trans, root,
2113                                                 parent, ref_root,
2114                                                 extent_op->flags_to_set,
2115                                                 &extent_op->key,
2116                                                 ref->level, &ins);
2117         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2118                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2119                                              node->num_bytes, parent, ref_root,
2120                                              ref->level, 0, 1, extent_op);
2121         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2122                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2123                                           node->num_bytes, parent, ref_root,
2124                                           ref->level, 0, 1, extent_op);
2125         } else {
2126                 BUG();
2127         }
2128         return ret;
2129 }
2130
2131 /* helper function to actually process a single delayed ref entry */
2132 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2133                                struct btrfs_root *root,
2134                                struct btrfs_delayed_ref_node *node,
2135                                struct btrfs_delayed_extent_op *extent_op,
2136                                int insert_reserved)
2137 {
2138         int ret = 0;
2139
2140         if (trans->aborted)
2141                 return 0;
2142
2143         if (btrfs_delayed_ref_is_head(node)) {
2144                 struct btrfs_delayed_ref_head *head;
2145                 /*
2146                  * we've hit the end of the chain and we were supposed
2147                  * to insert this extent into the tree.  But, it got
2148                  * deleted before we ever needed to insert it, so all
2149                  * we have to do is clean up the accounting
2150                  */
2151                 BUG_ON(extent_op);
2152                 head = btrfs_delayed_node_to_head(node);
2153                 if (insert_reserved) {
2154                         btrfs_pin_extent(root, node->bytenr,
2155                                          node->num_bytes, 1);
2156                         if (head->is_data) {
2157                                 ret = btrfs_del_csums(trans, root,
2158                                                       node->bytenr,
2159                                                       node->num_bytes);
2160                         }
2161                 }
2162                 mutex_unlock(&head->mutex);
2163                 return ret;
2164         }
2165
2166         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2167             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2168                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2169                                            insert_reserved);
2170         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2171                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2172                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2173                                            insert_reserved);
2174         else
2175                 BUG();
2176         return ret;
2177 }
2178
2179 static noinline struct btrfs_delayed_ref_node *
2180 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2181 {
2182         struct rb_node *node;
2183         struct btrfs_delayed_ref_node *ref;
2184         int action = BTRFS_ADD_DELAYED_REF;
2185 again:
2186         /*
2187          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2188          * this prevents ref count from going down to zero when
2189          * there still are pending delayed ref.
2190          */
2191         node = rb_prev(&head->node.rb_node);
2192         while (1) {
2193                 if (!node)
2194                         break;
2195                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2196                                 rb_node);
2197                 if (ref->bytenr != head->node.bytenr)
2198                         break;
2199                 if (ref->action == action)
2200                         return ref;
2201                 node = rb_prev(node);
2202         }
2203         if (action == BTRFS_ADD_DELAYED_REF) {
2204                 action = BTRFS_DROP_DELAYED_REF;
2205                 goto again;
2206         }
2207         return NULL;
2208 }
2209
2210 /*
2211  * Returns 0 on success or if called with an already aborted transaction.
2212  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2213  */
2214 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2215                                        struct btrfs_root *root,
2216                                        struct list_head *cluster)
2217 {
2218         struct btrfs_delayed_ref_root *delayed_refs;
2219         struct btrfs_delayed_ref_node *ref;
2220         struct btrfs_delayed_ref_head *locked_ref = NULL;
2221         struct btrfs_delayed_extent_op *extent_op;
2222         struct btrfs_fs_info *fs_info = root->fs_info;
2223         int ret;
2224         int count = 0;
2225         int must_insert_reserved = 0;
2226
2227         delayed_refs = &trans->transaction->delayed_refs;
2228         while (1) {
2229                 if (!locked_ref) {
2230                         /* pick a new head ref from the cluster list */
2231                         if (list_empty(cluster))
2232                                 break;
2233
2234                         locked_ref = list_entry(cluster->next,
2235                                      struct btrfs_delayed_ref_head, cluster);
2236
2237                         /* grab the lock that says we are going to process
2238                          * all the refs for this head */
2239                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2240
2241                         /*
2242                          * we may have dropped the spin lock to get the head
2243                          * mutex lock, and that might have given someone else
2244                          * time to free the head.  If that's true, it has been
2245                          * removed from our list and we can move on.
2246                          */
2247                         if (ret == -EAGAIN) {
2248                                 locked_ref = NULL;
2249                                 count++;
2250                                 continue;
2251                         }
2252                 }
2253
2254                 /*
2255                  * locked_ref is the head node, so we have to go one
2256                  * node back for any delayed ref updates
2257                  */
2258                 ref = select_delayed_ref(locked_ref);
2259
2260                 if (ref && ref->seq &&
2261                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2262                         /*
2263                          * there are still refs with lower seq numbers in the
2264                          * process of being added. Don't run this ref yet.
2265                          */
2266                         list_del_init(&locked_ref->cluster);
2267                         mutex_unlock(&locked_ref->mutex);
2268                         locked_ref = NULL;
2269                         delayed_refs->num_heads_ready++;
2270                         spin_unlock(&delayed_refs->lock);
2271                         cond_resched();
2272                         spin_lock(&delayed_refs->lock);
2273                         continue;
2274                 }
2275
2276                 /*
2277                  * record the must insert reserved flag before we
2278                  * drop the spin lock.
2279                  */
2280                 must_insert_reserved = locked_ref->must_insert_reserved;
2281                 locked_ref->must_insert_reserved = 0;
2282
2283                 extent_op = locked_ref->extent_op;
2284                 locked_ref->extent_op = NULL;
2285
2286                 if (!ref) {
2287                         /* All delayed refs have been processed, Go ahead
2288                          * and send the head node to run_one_delayed_ref,
2289                          * so that any accounting fixes can happen
2290                          */
2291                         ref = &locked_ref->node;
2292
2293                         if (extent_op && must_insert_reserved) {
2294                                 kfree(extent_op);
2295                                 extent_op = NULL;
2296                         }
2297
2298                         if (extent_op) {
2299                                 spin_unlock(&delayed_refs->lock);
2300
2301                                 ret = run_delayed_extent_op(trans, root,
2302                                                             ref, extent_op);
2303                                 kfree(extent_op);
2304
2305                                 if (ret) {
2306                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2307                                         spin_lock(&delayed_refs->lock);
2308                                         return ret;
2309                                 }
2310
2311                                 goto next;
2312                         }
2313
2314                         list_del_init(&locked_ref->cluster);
2315                         locked_ref = NULL;
2316                 }
2317
2318                 ref->in_tree = 0;
2319                 rb_erase(&ref->rb_node, &delayed_refs->root);
2320                 delayed_refs->num_entries--;
2321                 /*
2322                  * we modified num_entries, but as we're currently running
2323                  * delayed refs, skip
2324                  *     wake_up(&delayed_refs->seq_wait);
2325                  * here.
2326                  */
2327                 spin_unlock(&delayed_refs->lock);
2328
2329                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2330                                           must_insert_reserved);
2331
2332                 btrfs_put_delayed_ref(ref);
2333                 kfree(extent_op);
2334                 count++;
2335
2336                 if (ret) {
2337                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2338                         spin_lock(&delayed_refs->lock);
2339                         return ret;
2340                 }
2341
2342 next:
2343                 do_chunk_alloc(trans, fs_info->extent_root,
2344                                2 * 1024 * 1024,
2345                                btrfs_get_alloc_profile(root, 0),
2346                                CHUNK_ALLOC_NO_FORCE);
2347                 cond_resched();
2348                 spin_lock(&delayed_refs->lock);
2349         }
2350         return count;
2351 }
2352
2353 static void wait_for_more_refs(struct btrfs_fs_info *fs_info,
2354                                struct btrfs_delayed_ref_root *delayed_refs,
2355                                unsigned long num_refs,
2356                                struct list_head *first_seq)
2357 {
2358         spin_unlock(&delayed_refs->lock);
2359         pr_debug("waiting for more refs (num %ld, first %p)\n",
2360                  num_refs, first_seq);
2361         wait_event(fs_info->tree_mod_seq_wait,
2362                    num_refs != delayed_refs->num_entries ||
2363                    fs_info->tree_mod_seq_list.next != first_seq);
2364         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2365                  delayed_refs->num_entries, fs_info->tree_mod_seq_list.next);
2366         spin_lock(&delayed_refs->lock);
2367 }
2368
2369 #ifdef SCRAMBLE_DELAYED_REFS
2370 /*
2371  * Normally delayed refs get processed in ascending bytenr order. This
2372  * correlates in most cases to the order added. To expose dependencies on this
2373  * order, we start to process the tree in the middle instead of the beginning
2374  */
2375 static u64 find_middle(struct rb_root *root)
2376 {
2377         struct rb_node *n = root->rb_node;
2378         struct btrfs_delayed_ref_node *entry;
2379         int alt = 1;
2380         u64 middle;
2381         u64 first = 0, last = 0;
2382
2383         n = rb_first(root);
2384         if (n) {
2385                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2386                 first = entry->bytenr;
2387         }
2388         n = rb_last(root);
2389         if (n) {
2390                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2391                 last = entry->bytenr;
2392         }
2393         n = root->rb_node;
2394
2395         while (n) {
2396                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2397                 WARN_ON(!entry->in_tree);
2398
2399                 middle = entry->bytenr;
2400
2401                 if (alt)
2402                         n = n->rb_left;
2403                 else
2404                         n = n->rb_right;
2405
2406                 alt = 1 - alt;
2407         }
2408         return middle;
2409 }
2410 #endif
2411
2412 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2413                                          struct btrfs_fs_info *fs_info)
2414 {
2415         struct qgroup_update *qgroup_update;
2416         int ret = 0;
2417
2418         if (list_empty(&trans->qgroup_ref_list) !=
2419             !trans->delayed_ref_elem.seq) {
2420                 /* list without seq or seq without list */
2421                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2422                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2423                         trans->delayed_ref_elem.seq);
2424                 BUG();
2425         }
2426
2427         if (!trans->delayed_ref_elem.seq)
2428                 return 0;
2429
2430         while (!list_empty(&trans->qgroup_ref_list)) {
2431                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2432                                                  struct qgroup_update, list);
2433                 list_del(&qgroup_update->list);
2434                 if (!ret)
2435                         ret = btrfs_qgroup_account_ref(
2436                                         trans, fs_info, qgroup_update->node,
2437                                         qgroup_update->extent_op);
2438                 kfree(qgroup_update);
2439         }
2440
2441         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2442
2443         return ret;
2444 }
2445
2446 /*
2447  * this starts processing the delayed reference count updates and
2448  * extent insertions we have queued up so far.  count can be
2449  * 0, which means to process everything in the tree at the start
2450  * of the run (but not newly added entries), or it can be some target
2451  * number you'd like to process.
2452  *
2453  * Returns 0 on success or if called with an aborted transaction
2454  * Returns <0 on error and aborts the transaction
2455  */
2456 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2457                            struct btrfs_root *root, unsigned long count)
2458 {
2459         struct rb_node *node;
2460         struct btrfs_delayed_ref_root *delayed_refs;
2461         struct btrfs_delayed_ref_node *ref;
2462         struct list_head cluster;
2463         struct list_head *first_seq = NULL;
2464         int ret;
2465         u64 delayed_start;
2466         int run_all = count == (unsigned long)-1;
2467         int run_most = 0;
2468         unsigned long num_refs = 0;
2469         int consider_waiting;
2470
2471         /* We'll clean this up in btrfs_cleanup_transaction */
2472         if (trans->aborted)
2473                 return 0;
2474
2475         if (root == root->fs_info->extent_root)
2476                 root = root->fs_info->tree_root;
2477
2478         do_chunk_alloc(trans, root->fs_info->extent_root,
2479                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2480                        CHUNK_ALLOC_NO_FORCE);
2481
2482         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2483
2484         delayed_refs = &trans->transaction->delayed_refs;
2485         INIT_LIST_HEAD(&cluster);
2486 again:
2487         consider_waiting = 0;
2488         spin_lock(&delayed_refs->lock);
2489
2490 #ifdef SCRAMBLE_DELAYED_REFS
2491         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2492 #endif
2493
2494         if (count == 0) {
2495                 count = delayed_refs->num_entries * 2;
2496                 run_most = 1;
2497         }
2498         while (1) {
2499                 if (!(run_all || run_most) &&
2500                     delayed_refs->num_heads_ready < 64)
2501                         break;
2502
2503                 /*
2504                  * go find something we can process in the rbtree.  We start at
2505                  * the beginning of the tree, and then build a cluster
2506                  * of refs to process starting at the first one we are able to
2507                  * lock
2508                  */
2509                 delayed_start = delayed_refs->run_delayed_start;
2510                 ret = btrfs_find_ref_cluster(trans, &cluster,
2511                                              delayed_refs->run_delayed_start);
2512                 if (ret)
2513                         break;
2514
2515                 if (delayed_start >= delayed_refs->run_delayed_start) {
2516                         if (consider_waiting == 0) {
2517                                 /*
2518                                  * btrfs_find_ref_cluster looped. let's do one
2519                                  * more cycle. if we don't run any delayed ref
2520                                  * during that cycle (because we can't because
2521                                  * all of them are blocked) and if the number of
2522                                  * refs doesn't change, we avoid busy waiting.
2523                                  */
2524                                 consider_waiting = 1;
2525                                 num_refs = delayed_refs->num_entries;
2526                                 first_seq = root->fs_info->tree_mod_seq_list.next;
2527                         } else {
2528                                 wait_for_more_refs(root->fs_info, delayed_refs,
2529                                                    num_refs, first_seq);
2530                                 /*
2531                                  * after waiting, things have changed. we
2532                                  * dropped the lock and someone else might have
2533                                  * run some refs, built new clusters and so on.
2534                                  * therefore, we restart staleness detection.
2535                                  */
2536                                 consider_waiting = 0;
2537                         }
2538                 }
2539
2540                 ret = run_clustered_refs(trans, root, &cluster);
2541                 if (ret < 0) {
2542                         spin_unlock(&delayed_refs->lock);
2543                         btrfs_abort_transaction(trans, root, ret);
2544                         return ret;
2545                 }
2546
2547                 count -= min_t(unsigned long, ret, count);
2548
2549                 if (count == 0)
2550                         break;
2551
2552                 if (ret || delayed_refs->run_delayed_start == 0) {
2553                         /* refs were run, let's reset staleness detection */
2554                         consider_waiting = 0;
2555                 }
2556         }
2557
2558         if (run_all) {
2559                 node = rb_first(&delayed_refs->root);
2560                 if (!node)
2561                         goto out;
2562                 count = (unsigned long)-1;
2563
2564                 while (node) {
2565                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2566                                        rb_node);
2567                         if (btrfs_delayed_ref_is_head(ref)) {
2568                                 struct btrfs_delayed_ref_head *head;
2569
2570                                 head = btrfs_delayed_node_to_head(ref);
2571                                 atomic_inc(&ref->refs);
2572
2573                                 spin_unlock(&delayed_refs->lock);
2574                                 /*
2575                                  * Mutex was contended, block until it's
2576                                  * released and try again
2577                                  */
2578                                 mutex_lock(&head->mutex);
2579                                 mutex_unlock(&head->mutex);
2580
2581                                 btrfs_put_delayed_ref(ref);
2582                                 cond_resched();
2583                                 goto again;
2584                         }
2585                         node = rb_next(node);
2586                 }
2587                 spin_unlock(&delayed_refs->lock);
2588                 schedule_timeout(1);
2589                 goto again;
2590         }
2591 out:
2592         spin_unlock(&delayed_refs->lock);
2593         assert_qgroups_uptodate(trans);
2594         return 0;
2595 }
2596
2597 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2598                                 struct btrfs_root *root,
2599                                 u64 bytenr, u64 num_bytes, u64 flags,
2600                                 int is_data)
2601 {
2602         struct btrfs_delayed_extent_op *extent_op;
2603         int ret;
2604
2605         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2606         if (!extent_op)
2607                 return -ENOMEM;
2608
2609         extent_op->flags_to_set = flags;
2610         extent_op->update_flags = 1;
2611         extent_op->update_key = 0;
2612         extent_op->is_data = is_data ? 1 : 0;
2613
2614         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2615                                           num_bytes, extent_op);
2616         if (ret)
2617                 kfree(extent_op);
2618         return ret;
2619 }
2620
2621 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2622                                       struct btrfs_root *root,
2623                                       struct btrfs_path *path,
2624                                       u64 objectid, u64 offset, u64 bytenr)
2625 {
2626         struct btrfs_delayed_ref_head *head;
2627         struct btrfs_delayed_ref_node *ref;
2628         struct btrfs_delayed_data_ref *data_ref;
2629         struct btrfs_delayed_ref_root *delayed_refs;
2630         struct rb_node *node;
2631         int ret = 0;
2632
2633         ret = -ENOENT;
2634         delayed_refs = &trans->transaction->delayed_refs;
2635         spin_lock(&delayed_refs->lock);
2636         head = btrfs_find_delayed_ref_head(trans, bytenr);
2637         if (!head)
2638                 goto out;
2639
2640         if (!mutex_trylock(&head->mutex)) {
2641                 atomic_inc(&head->node.refs);
2642                 spin_unlock(&delayed_refs->lock);
2643
2644                 btrfs_release_path(path);
2645
2646                 /*
2647                  * Mutex was contended, block until it's released and let
2648                  * caller try again
2649                  */
2650                 mutex_lock(&head->mutex);
2651                 mutex_unlock(&head->mutex);
2652                 btrfs_put_delayed_ref(&head->node);
2653                 return -EAGAIN;
2654         }
2655
2656         node = rb_prev(&head->node.rb_node);
2657         if (!node)
2658                 goto out_unlock;
2659
2660         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2661
2662         if (ref->bytenr != bytenr)
2663                 goto out_unlock;
2664
2665         ret = 1;
2666         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2667                 goto out_unlock;
2668
2669         data_ref = btrfs_delayed_node_to_data_ref(ref);
2670
2671         node = rb_prev(node);
2672         if (node) {
2673                 int seq = ref->seq;
2674
2675                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2676                 if (ref->bytenr == bytenr && ref->seq == seq)
2677                         goto out_unlock;
2678         }
2679
2680         if (data_ref->root != root->root_key.objectid ||
2681             data_ref->objectid != objectid || data_ref->offset != offset)
2682                 goto out_unlock;
2683
2684         ret = 0;
2685 out_unlock:
2686         mutex_unlock(&head->mutex);
2687 out:
2688         spin_unlock(&delayed_refs->lock);
2689         return ret;
2690 }
2691
2692 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2693                                         struct btrfs_root *root,
2694                                         struct btrfs_path *path,
2695                                         u64 objectid, u64 offset, u64 bytenr)
2696 {
2697         struct btrfs_root *extent_root = root->fs_info->extent_root;
2698         struct extent_buffer *leaf;
2699         struct btrfs_extent_data_ref *ref;
2700         struct btrfs_extent_inline_ref *iref;
2701         struct btrfs_extent_item *ei;
2702         struct btrfs_key key;
2703         u32 item_size;
2704         int ret;
2705
2706         key.objectid = bytenr;
2707         key.offset = (u64)-1;
2708         key.type = BTRFS_EXTENT_ITEM_KEY;
2709
2710         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2711         if (ret < 0)
2712                 goto out;
2713         BUG_ON(ret == 0); /* Corruption */
2714
2715         ret = -ENOENT;
2716         if (path->slots[0] == 0)
2717                 goto out;
2718
2719         path->slots[0]--;
2720         leaf = path->nodes[0];
2721         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2722
2723         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2724                 goto out;
2725
2726         ret = 1;
2727         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2728 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2729         if (item_size < sizeof(*ei)) {
2730                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2731                 goto out;
2732         }
2733 #endif
2734         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2735
2736         if (item_size != sizeof(*ei) +
2737             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2738                 goto out;
2739
2740         if (btrfs_extent_generation(leaf, ei) <=
2741             btrfs_root_last_snapshot(&root->root_item))
2742                 goto out;
2743
2744         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2745         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2746             BTRFS_EXTENT_DATA_REF_KEY)
2747                 goto out;
2748
2749         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2750         if (btrfs_extent_refs(leaf, ei) !=
2751             btrfs_extent_data_ref_count(leaf, ref) ||
2752             btrfs_extent_data_ref_root(leaf, ref) !=
2753             root->root_key.objectid ||
2754             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2755             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2756                 goto out;
2757
2758         ret = 0;
2759 out:
2760         return ret;
2761 }
2762
2763 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2764                           struct btrfs_root *root,
2765                           u64 objectid, u64 offset, u64 bytenr)
2766 {
2767         struct btrfs_path *path;
2768         int ret;
2769         int ret2;
2770
2771         path = btrfs_alloc_path();
2772         if (!path)
2773                 return -ENOENT;
2774
2775         do {
2776                 ret = check_committed_ref(trans, root, path, objectid,
2777                                           offset, bytenr);
2778                 if (ret && ret != -ENOENT)
2779                         goto out;
2780
2781                 ret2 = check_delayed_ref(trans, root, path, objectid,
2782                                          offset, bytenr);
2783         } while (ret2 == -EAGAIN);
2784
2785         if (ret2 && ret2 != -ENOENT) {
2786                 ret = ret2;
2787                 goto out;
2788         }
2789
2790         if (ret != -ENOENT || ret2 != -ENOENT)
2791                 ret = 0;
2792 out:
2793         btrfs_free_path(path);
2794         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2795                 WARN_ON(ret > 0);
2796         return ret;
2797 }
2798
2799 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2800                            struct btrfs_root *root,
2801                            struct extent_buffer *buf,
2802                            int full_backref, int inc, int for_cow)
2803 {
2804         u64 bytenr;
2805         u64 num_bytes;
2806         u64 parent;
2807         u64 ref_root;
2808         u32 nritems;
2809         struct btrfs_key key;
2810         struct btrfs_file_extent_item *fi;
2811         int i;
2812         int level;
2813         int ret = 0;
2814         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2815                             u64, u64, u64, u64, u64, u64, int);
2816
2817         ref_root = btrfs_header_owner(buf);
2818         nritems = btrfs_header_nritems(buf);
2819         level = btrfs_header_level(buf);
2820
2821         if (!root->ref_cows && level == 0)
2822                 return 0;
2823
2824         if (inc)
2825                 process_func = btrfs_inc_extent_ref;
2826         else
2827                 process_func = btrfs_free_extent;
2828
2829         if (full_backref)
2830                 parent = buf->start;
2831         else
2832                 parent = 0;
2833
2834         for (i = 0; i < nritems; i++) {
2835                 if (level == 0) {
2836                         btrfs_item_key_to_cpu(buf, &key, i);
2837                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2838                                 continue;
2839                         fi = btrfs_item_ptr(buf, i,
2840                                             struct btrfs_file_extent_item);
2841                         if (btrfs_file_extent_type(buf, fi) ==
2842                             BTRFS_FILE_EXTENT_INLINE)
2843                                 continue;
2844                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2845                         if (bytenr == 0)
2846                                 continue;
2847
2848                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2849                         key.offset -= btrfs_file_extent_offset(buf, fi);
2850                         ret = process_func(trans, root, bytenr, num_bytes,
2851                                            parent, ref_root, key.objectid,
2852                                            key.offset, for_cow);
2853                         if (ret)
2854                                 goto fail;
2855                 } else {
2856                         bytenr = btrfs_node_blockptr(buf, i);
2857                         num_bytes = btrfs_level_size(root, level - 1);
2858                         ret = process_func(trans, root, bytenr, num_bytes,
2859                                            parent, ref_root, level - 1, 0,
2860                                            for_cow);
2861                         if (ret)
2862                                 goto fail;
2863                 }
2864         }
2865         return 0;
2866 fail:
2867         return ret;
2868 }
2869
2870 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2871                   struct extent_buffer *buf, int full_backref, int for_cow)
2872 {
2873         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2874 }
2875
2876 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2877                   struct extent_buffer *buf, int full_backref, int for_cow)
2878 {
2879         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2880 }
2881
2882 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2883                                  struct btrfs_root *root,
2884                                  struct btrfs_path *path,
2885                                  struct btrfs_block_group_cache *cache)
2886 {
2887         int ret;
2888         struct btrfs_root *extent_root = root->fs_info->extent_root;
2889         unsigned long bi;
2890         struct extent_buffer *leaf;
2891
2892         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2893         if (ret < 0)
2894                 goto fail;
2895         BUG_ON(ret); /* Corruption */
2896
2897         leaf = path->nodes[0];
2898         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2899         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2900         btrfs_mark_buffer_dirty(leaf);
2901         btrfs_release_path(path);
2902 fail:
2903         if (ret) {
2904                 btrfs_abort_transaction(trans, root, ret);
2905                 return ret;
2906         }
2907         return 0;
2908
2909 }
2910
2911 static struct btrfs_block_group_cache *
2912 next_block_group(struct btrfs_root *root,
2913                  struct btrfs_block_group_cache *cache)
2914 {
2915         struct rb_node *node;
2916         spin_lock(&root->fs_info->block_group_cache_lock);
2917         node = rb_next(&cache->cache_node);
2918         btrfs_put_block_group(cache);
2919         if (node) {
2920                 cache = rb_entry(node, struct btrfs_block_group_cache,
2921                                  cache_node);
2922                 btrfs_get_block_group(cache);
2923         } else
2924                 cache = NULL;
2925         spin_unlock(&root->fs_info->block_group_cache_lock);
2926         return cache;
2927 }
2928
2929 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2930                             struct btrfs_trans_handle *trans,
2931                             struct btrfs_path *path)
2932 {
2933         struct btrfs_root *root = block_group->fs_info->tree_root;
2934         struct inode *inode = NULL;
2935         u64 alloc_hint = 0;
2936         int dcs = BTRFS_DC_ERROR;
2937         int num_pages = 0;
2938         int retries = 0;
2939         int ret = 0;
2940
2941         /*
2942          * If this block group is smaller than 100 megs don't bother caching the
2943          * block group.
2944          */
2945         if (block_group->key.offset < (100 * 1024 * 1024)) {
2946                 spin_lock(&block_group->lock);
2947                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2948                 spin_unlock(&block_group->lock);
2949                 return 0;
2950         }
2951
2952 again:
2953         inode = lookup_free_space_inode(root, block_group, path);
2954         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2955                 ret = PTR_ERR(inode);
2956                 btrfs_release_path(path);
2957                 goto out;
2958         }
2959
2960         if (IS_ERR(inode)) {
2961                 BUG_ON(retries);
2962                 retries++;
2963
2964                 if (block_group->ro)
2965                         goto out_free;
2966
2967                 ret = create_free_space_inode(root, trans, block_group, path);
2968                 if (ret)
2969                         goto out_free;
2970                 goto again;
2971         }
2972
2973         /* We've already setup this transaction, go ahead and exit */
2974         if (block_group->cache_generation == trans->transid &&
2975             i_size_read(inode)) {
2976                 dcs = BTRFS_DC_SETUP;
2977                 goto out_put;
2978         }
2979
2980         /*
2981          * We want to set the generation to 0, that way if anything goes wrong
2982          * from here on out we know not to trust this cache when we load up next
2983          * time.
2984          */
2985         BTRFS_I(inode)->generation = 0;
2986         ret = btrfs_update_inode(trans, root, inode);
2987         WARN_ON(ret);
2988
2989         if (i_size_read(inode) > 0) {
2990                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2991                                                       inode);
2992                 if (ret)
2993                         goto out_put;
2994         }
2995
2996         spin_lock(&block_group->lock);
2997         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2998             !btrfs_test_opt(root, SPACE_CACHE)) {
2999                 /*
3000                  * don't bother trying to write stuff out _if_
3001                  * a) we're not cached,
3002                  * b) we're with nospace_cache mount option.
3003                  */
3004                 dcs = BTRFS_DC_WRITTEN;
3005                 spin_unlock(&block_group->lock);
3006                 goto out_put;
3007         }
3008         spin_unlock(&block_group->lock);
3009
3010         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
3011         if (!num_pages)
3012                 num_pages = 1;
3013
3014         /*
3015          * Just to make absolutely sure we have enough space, we're going to
3016          * preallocate 12 pages worth of space for each block group.  In
3017          * practice we ought to use at most 8, but we need extra space so we can
3018          * add our header and have a terminator between the extents and the
3019          * bitmaps.
3020          */
3021         num_pages *= 16;
3022         num_pages *= PAGE_CACHE_SIZE;
3023
3024         ret = btrfs_check_data_free_space(inode, num_pages);
3025         if (ret)
3026                 goto out_put;
3027
3028         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3029                                               num_pages, num_pages,
3030                                               &alloc_hint);
3031         if (!ret)
3032                 dcs = BTRFS_DC_SETUP;
3033         btrfs_free_reserved_data_space(inode, num_pages);
3034
3035 out_put:
3036         iput(inode);
3037 out_free:
3038         btrfs_release_path(path);
3039 out:
3040         spin_lock(&block_group->lock);
3041         if (!ret && dcs == BTRFS_DC_SETUP)
3042                 block_group->cache_generation = trans->transid;
3043         block_group->disk_cache_state = dcs;
3044         spin_unlock(&block_group->lock);
3045
3046         return ret;
3047 }
3048
3049 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3050                                    struct btrfs_root *root)
3051 {
3052         struct btrfs_block_group_cache *cache;
3053         int err = 0;
3054         struct btrfs_path *path;
3055         u64 last = 0;
3056
3057         path = btrfs_alloc_path();
3058         if (!path)
3059                 return -ENOMEM;
3060
3061 again:
3062         while (1) {
3063                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3064                 while (cache) {
3065                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3066                                 break;
3067                         cache = next_block_group(root, cache);
3068                 }
3069                 if (!cache) {
3070                         if (last == 0)
3071                                 break;
3072                         last = 0;
3073                         continue;
3074                 }
3075                 err = cache_save_setup(cache, trans, path);
3076                 last = cache->key.objectid + cache->key.offset;
3077                 btrfs_put_block_group(cache);
3078         }
3079
3080         while (1) {
3081                 if (last == 0) {
3082                         err = btrfs_run_delayed_refs(trans, root,
3083                                                      (unsigned long)-1);
3084                         if (err) /* File system offline */
3085                                 goto out;
3086                 }
3087
3088                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3089                 while (cache) {
3090                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3091                                 btrfs_put_block_group(cache);
3092                                 goto again;
3093                         }
3094
3095                         if (cache->dirty)
3096                                 break;
3097                         cache = next_block_group(root, cache);
3098                 }
3099                 if (!cache) {
3100                         if (last == 0)
3101                                 break;
3102                         last = 0;
3103                         continue;
3104                 }
3105
3106                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3107                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3108                 cache->dirty = 0;
3109                 last = cache->key.objectid + cache->key.offset;
3110
3111                 err = write_one_cache_group(trans, root, path, cache);
3112                 if (err) /* File system offline */
3113                         goto out;
3114
3115                 btrfs_put_block_group(cache);
3116         }
3117
3118         while (1) {
3119                 /*
3120                  * I don't think this is needed since we're just marking our
3121                  * preallocated extent as written, but just in case it can't
3122                  * hurt.
3123                  */
3124                 if (last == 0) {
3125                         err = btrfs_run_delayed_refs(trans, root,
3126                                                      (unsigned long)-1);
3127                         if (err) /* File system offline */
3128                                 goto out;
3129                 }
3130
3131                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3132                 while (cache) {
3133                         /*
3134                          * Really this shouldn't happen, but it could if we
3135                          * couldn't write the entire preallocated extent and
3136                          * splitting the extent resulted in a new block.
3137                          */
3138                         if (cache->dirty) {
3139                                 btrfs_put_block_group(cache);
3140                                 goto again;
3141                         }
3142                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3143                                 break;
3144                         cache = next_block_group(root, cache);
3145                 }
3146                 if (!cache) {
3147                         if (last == 0)
3148                                 break;
3149                         last = 0;
3150                         continue;
3151                 }
3152
3153                 err = btrfs_write_out_cache(root, trans, cache, path);
3154
3155                 /*
3156                  * If we didn't have an error then the cache state is still
3157                  * NEED_WRITE, so we can set it to WRITTEN.
3158                  */
3159                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3160                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3161                 last = cache->key.objectid + cache->key.offset;
3162                 btrfs_put_block_group(cache);
3163         }
3164 out:
3165
3166         btrfs_free_path(path);
3167         return err;
3168 }
3169
3170 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3171 {
3172         struct btrfs_block_group_cache *block_group;
3173         int readonly = 0;
3174
3175         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3176         if (!block_group || block_group->ro)
3177                 readonly = 1;
3178         if (block_group)
3179                 btrfs_put_block_group(block_group);
3180         return readonly;
3181 }
3182
3183 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3184                              u64 total_bytes, u64 bytes_used,
3185                              struct btrfs_space_info **space_info)
3186 {
3187         struct btrfs_space_info *found;
3188         int i;
3189         int factor;
3190
3191         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3192                      BTRFS_BLOCK_GROUP_RAID10))
3193                 factor = 2;
3194         else
3195                 factor = 1;
3196
3197         found = __find_space_info(info, flags);
3198         if (found) {
3199                 spin_lock(&found->lock);
3200                 found->total_bytes += total_bytes;
3201                 found->disk_total += total_bytes * factor;
3202                 found->bytes_used += bytes_used;
3203                 found->disk_used += bytes_used * factor;
3204                 found->full = 0;
3205                 spin_unlock(&found->lock);
3206                 *space_info = found;
3207                 return 0;
3208         }
3209         found = kzalloc(sizeof(*found), GFP_NOFS);
3210         if (!found)
3211                 return -ENOMEM;
3212
3213         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3214                 INIT_LIST_HEAD(&found->block_groups[i]);
3215         init_rwsem(&found->groups_sem);
3216         spin_lock_init(&found->lock);
3217         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3218         found->total_bytes = total_bytes;
3219         found->disk_total = total_bytes * factor;
3220         found->bytes_used = bytes_used;
3221         found->disk_used = bytes_used * factor;
3222         found->bytes_pinned = 0;
3223         found->bytes_reserved = 0;
3224         found->bytes_readonly = 0;
3225         found->bytes_may_use = 0;
3226         found->full = 0;
3227         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3228         found->chunk_alloc = 0;
3229         found->flush = 0;
3230         init_waitqueue_head(&found->wait);
3231         *space_info = found;
3232         list_add_rcu(&found->list, &info->space_info);
3233         if (flags & BTRFS_BLOCK_GROUP_DATA)
3234                 info->data_sinfo = found;
3235         return 0;
3236 }
3237
3238 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3239 {
3240         u64 extra_flags = chunk_to_extended(flags) &
3241                                 BTRFS_EXTENDED_PROFILE_MASK;
3242
3243         if (flags & BTRFS_BLOCK_GROUP_DATA)
3244                 fs_info->avail_data_alloc_bits |= extra_flags;
3245         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3246                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3247         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3248                 fs_info->avail_system_alloc_bits |= extra_flags;
3249 }
3250
3251 /*
3252  * returns target flags in extended format or 0 if restripe for this
3253  * chunk_type is not in progress
3254  *
3255  * should be called with either volume_mutex or balance_lock held
3256  */
3257 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3258 {
3259         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3260         u64 target = 0;
3261
3262         if (!bctl)
3263                 return 0;
3264
3265         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3266             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3267                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3268         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3269                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3270                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3271         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3272                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3273                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3274         }
3275
3276         return target;
3277 }
3278
3279 /*
3280  * @flags: available profiles in extended format (see ctree.h)
3281  *
3282  * Returns reduced profile in chunk format.  If profile changing is in
3283  * progress (either running or paused) picks the target profile (if it's
3284  * already available), otherwise falls back to plain reducing.
3285  */
3286 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3287 {
3288         /*
3289          * we add in the count of missing devices because we want
3290          * to make sure that any RAID levels on a degraded FS
3291          * continue to be honored.
3292          */
3293         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3294                 root->fs_info->fs_devices->missing_devices;
3295         u64 target;
3296
3297         /*
3298          * see if restripe for this chunk_type is in progress, if so
3299          * try to reduce to the target profile
3300          */
3301         spin_lock(&root->fs_info->balance_lock);
3302         target = get_restripe_target(root->fs_info, flags);
3303         if (target) {
3304                 /* pick target profile only if it's already available */
3305                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3306                         spin_unlock(&root->fs_info->balance_lock);
3307                         return extended_to_chunk(target);
3308                 }
3309         }
3310         spin_unlock(&root->fs_info->balance_lock);
3311
3312         if (num_devices == 1)
3313                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3314         if (num_devices < 4)
3315                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3316
3317         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3318             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3319                       BTRFS_BLOCK_GROUP_RAID10))) {
3320                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3321         }
3322
3323         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3324             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3325                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3326         }
3327
3328         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3329             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3330              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3331              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3332                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3333         }
3334
3335         return extended_to_chunk(flags);
3336 }
3337
3338 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3339 {
3340         if (flags & BTRFS_BLOCK_GROUP_DATA)
3341                 flags |= root->fs_info->avail_data_alloc_bits;
3342         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3343                 flags |= root->fs_info->avail_system_alloc_bits;
3344         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3345                 flags |= root->fs_info->avail_metadata_alloc_bits;
3346
3347         return btrfs_reduce_alloc_profile(root, flags);
3348 }
3349
3350 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3351 {
3352         u64 flags;
3353
3354         if (data)
3355                 flags = BTRFS_BLOCK_GROUP_DATA;
3356         else if (root == root->fs_info->chunk_root)
3357                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3358         else
3359                 flags = BTRFS_BLOCK_GROUP_METADATA;
3360
3361         return get_alloc_profile(root, flags);
3362 }
3363
3364 /*
3365  * This will check the space that the inode allocates from to make sure we have
3366  * enough space for bytes.
3367  */
3368 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3369 {
3370         struct btrfs_space_info *data_sinfo;
3371         struct btrfs_root *root = BTRFS_I(inode)->root;
3372         struct btrfs_fs_info *fs_info = root->fs_info;
3373         u64 used;
3374         int ret = 0, committed = 0, alloc_chunk = 1;
3375
3376         /* make sure bytes are sectorsize aligned */
3377         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3378
3379         if (root == root->fs_info->tree_root ||
3380             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3381                 alloc_chunk = 0;
3382                 committed = 1;
3383         }
3384
3385         data_sinfo = fs_info->data_sinfo;
3386         if (!data_sinfo)
3387                 goto alloc;
3388
3389 again:
3390         /* make sure we have enough space to handle the data first */
3391         spin_lock(&data_sinfo->lock);
3392         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3393                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3394                 data_sinfo->bytes_may_use;
3395
3396         if (used + bytes > data_sinfo->total_bytes) {
3397                 struct btrfs_trans_handle *trans;
3398
3399                 /*
3400                  * if we don't have enough free bytes in this space then we need
3401                  * to alloc a new chunk.
3402                  */
3403                 if (!data_sinfo->full && alloc_chunk) {
3404                         u64 alloc_target;
3405
3406                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3407                         spin_unlock(&data_sinfo->lock);
3408 alloc:
3409                         alloc_target = btrfs_get_alloc_profile(root, 1);
3410                         trans = btrfs_join_transaction(root);
3411                         if (IS_ERR(trans))
3412                                 return PTR_ERR(trans);
3413
3414                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3415                                              bytes + 2 * 1024 * 1024,
3416                                              alloc_target,
3417                                              CHUNK_ALLOC_NO_FORCE);
3418                         btrfs_end_transaction(trans, root);
3419                         if (ret < 0) {
3420                                 if (ret != -ENOSPC)
3421                                         return ret;
3422                                 else
3423                                         goto commit_trans;
3424                         }
3425
3426                         if (!data_sinfo)
3427                                 data_sinfo = fs_info->data_sinfo;
3428
3429                         goto again;
3430                 }
3431
3432                 /*
3433                  * If we have less pinned bytes than we want to allocate then
3434                  * don't bother committing the transaction, it won't help us.
3435                  */
3436                 if (data_sinfo->bytes_pinned < bytes)
3437                         committed = 1;
3438                 spin_unlock(&data_sinfo->lock);
3439
3440                 /* commit the current transaction and try again */
3441 commit_trans:
3442                 if (!committed &&
3443                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3444                         committed = 1;
3445                         trans = btrfs_join_transaction(root);
3446                         if (IS_ERR(trans))
3447                                 return PTR_ERR(trans);
3448                         ret = btrfs_commit_transaction(trans, root);
3449                         if (ret)
3450                                 return ret;
3451                         goto again;
3452                 }
3453
3454                 return -ENOSPC;
3455         }
3456         data_sinfo->bytes_may_use += bytes;
3457         trace_btrfs_space_reservation(root->fs_info, "space_info",
3458                                       data_sinfo->flags, bytes, 1);
3459         spin_unlock(&data_sinfo->lock);
3460
3461         return 0;
3462 }
3463
3464 /*
3465  * Called if we need to clear a data reservation for this inode.
3466  */
3467 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3468 {
3469         struct btrfs_root *root = BTRFS_I(inode)->root;
3470         struct btrfs_space_info *data_sinfo;
3471
3472         /* make sure bytes are sectorsize aligned */
3473         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3474
3475         data_sinfo = root->fs_info->data_sinfo;
3476         spin_lock(&data_sinfo->lock);
3477         data_sinfo->bytes_may_use -= bytes;
3478         trace_btrfs_space_reservation(root->fs_info, "space_info",
3479                                       data_sinfo->flags, bytes, 0);
3480         spin_unlock(&data_sinfo->lock);
3481 }
3482
3483 static void force_metadata_allocation(struct btrfs_fs_info *info)
3484 {
3485         struct list_head *head = &info->space_info;
3486         struct btrfs_space_info *found;
3487
3488         rcu_read_lock();
3489         list_for_each_entry_rcu(found, head, list) {
3490                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3491                         found->force_alloc = CHUNK_ALLOC_FORCE;
3492         }
3493         rcu_read_unlock();
3494 }
3495
3496 static int should_alloc_chunk(struct btrfs_root *root,
3497                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3498                               int force)
3499 {
3500         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3501         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3502         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3503         u64 thresh;
3504
3505         if (force == CHUNK_ALLOC_FORCE)
3506                 return 1;
3507
3508         /*
3509          * We need to take into account the global rsv because for all intents
3510          * and purposes it's used space.  Don't worry about locking the
3511          * global_rsv, it doesn't change except when the transaction commits.
3512          */
3513         num_allocated += global_rsv->size;
3514
3515         /*
3516          * in limited mode, we want to have some free space up to
3517          * about 1% of the FS size.
3518          */
3519         if (force == CHUNK_ALLOC_LIMITED) {
3520                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3521                 thresh = max_t(u64, 64 * 1024 * 1024,
3522                                div_factor_fine(thresh, 1));
3523
3524                 if (num_bytes - num_allocated < thresh)
3525                         return 1;
3526         }
3527         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3528
3529         /* 256MB or 2% of the FS */
3530         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3531         /* system chunks need a much small threshold */
3532         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3533                 thresh = 32 * 1024 * 1024;
3534
3535         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3536                 return 0;
3537         return 1;
3538 }
3539
3540 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3541 {
3542         u64 num_dev;
3543
3544         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3545             type & BTRFS_BLOCK_GROUP_RAID0)
3546                 num_dev = root->fs_info->fs_devices->rw_devices;
3547         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3548                 num_dev = 2;
3549         else
3550                 num_dev = 1;    /* DUP or single */
3551
3552         /* metadata for updaing devices and chunk tree */
3553         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3554 }
3555
3556 static void check_system_chunk(struct btrfs_trans_handle *trans,
3557                                struct btrfs_root *root, u64 type)
3558 {
3559         struct btrfs_space_info *info;
3560         u64 left;
3561         u64 thresh;
3562
3563         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3564         spin_lock(&info->lock);
3565         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3566                 info->bytes_reserved - info->bytes_readonly;
3567         spin_unlock(&info->lock);
3568
3569         thresh = get_system_chunk_thresh(root, type);
3570         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3571                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3572                        left, thresh, type);
3573                 dump_space_info(info, 0, 0);
3574         }
3575
3576         if (left < thresh) {
3577                 u64 flags;
3578
3579                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3580                 btrfs_alloc_chunk(trans, root, flags);
3581         }
3582 }
3583
3584 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3585                           struct btrfs_root *extent_root, u64 alloc_bytes,
3586                           u64 flags, int force)
3587 {
3588         struct btrfs_space_info *space_info;
3589         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3590         int wait_for_alloc = 0;
3591         int ret = 0;
3592
3593         space_info = __find_space_info(extent_root->fs_info, flags);
3594         if (!space_info) {
3595                 ret = update_space_info(extent_root->fs_info, flags,
3596                                         0, 0, &space_info);
3597                 BUG_ON(ret); /* -ENOMEM */
3598         }
3599         BUG_ON(!space_info); /* Logic error */
3600
3601 again:
3602         spin_lock(&space_info->lock);
3603         if (force < space_info->force_alloc)
3604                 force = space_info->force_alloc;
3605         if (space_info->full) {
3606                 spin_unlock(&space_info->lock);
3607                 return 0;
3608         }
3609
3610         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3611                 spin_unlock(&space_info->lock);
3612                 return 0;
3613         } else if (space_info->chunk_alloc) {
3614                 wait_for_alloc = 1;
3615         } else {
3616                 space_info->chunk_alloc = 1;
3617         }
3618
3619         spin_unlock(&space_info->lock);
3620
3621         mutex_lock(&fs_info->chunk_mutex);
3622
3623         /*
3624          * The chunk_mutex is held throughout the entirety of a chunk
3625          * allocation, so once we've acquired the chunk_mutex we know that the
3626          * other guy is done and we need to recheck and see if we should
3627          * allocate.
3628          */
3629         if (wait_for_alloc) {
3630                 mutex_unlock(&fs_info->chunk_mutex);
3631                 wait_for_alloc = 0;
3632                 goto again;
3633         }
3634
3635         /*
3636          * If we have mixed data/metadata chunks we want to make sure we keep
3637          * allocating mixed chunks instead of individual chunks.
3638          */
3639         if (btrfs_mixed_space_info(space_info))
3640                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3641
3642         /*
3643          * if we're doing a data chunk, go ahead and make sure that
3644          * we keep a reasonable number of metadata chunks allocated in the
3645          * FS as well.
3646          */
3647         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3648                 fs_info->data_chunk_allocations++;
3649                 if (!(fs_info->data_chunk_allocations %
3650                       fs_info->metadata_ratio))
3651                         force_metadata_allocation(fs_info);
3652         }
3653
3654         /*
3655          * Check if we have enough space in SYSTEM chunk because we may need
3656          * to update devices.
3657          */
3658         check_system_chunk(trans, extent_root, flags);
3659
3660         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3661         if (ret < 0 && ret != -ENOSPC)
3662                 goto out;
3663
3664         spin_lock(&space_info->lock);
3665         if (ret)
3666                 space_info->full = 1;
3667         else
3668                 ret = 1;
3669
3670         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3671         space_info->chunk_alloc = 0;
3672         spin_unlock(&space_info->lock);
3673 out:
3674         mutex_unlock(&fs_info->chunk_mutex);
3675         return ret;
3676 }
3677
3678 /*
3679  * shrink metadata reservation for delalloc
3680  */
3681 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3682                             bool wait_ordered)
3683 {
3684         struct btrfs_block_rsv *block_rsv;
3685         struct btrfs_space_info *space_info;
3686         struct btrfs_trans_handle *trans;
3687         u64 delalloc_bytes;
3688         u64 max_reclaim;
3689         long time_left;
3690         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3691         int loops = 0;
3692
3693         trans = (struct btrfs_trans_handle *)current->journal_info;
3694         block_rsv = &root->fs_info->delalloc_block_rsv;
3695         space_info = block_rsv->space_info;
3696
3697         smp_mb();
3698         delalloc_bytes = root->fs_info->delalloc_bytes;
3699         if (delalloc_bytes == 0) {
3700                 if (trans)
3701                         return;
3702                 btrfs_wait_ordered_extents(root, 0, 0);
3703                 return;
3704         }
3705
3706         while (delalloc_bytes && loops < 3) {
3707                 max_reclaim = min(delalloc_bytes, to_reclaim);
3708                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3709                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3710                                                WB_REASON_FS_FREE_SPACE);
3711
3712                 spin_lock(&space_info->lock);
3713                 if (space_info->bytes_used + space_info->bytes_reserved +
3714                     space_info->bytes_pinned + space_info->bytes_readonly +
3715                     space_info->bytes_may_use + orig <=
3716                     space_info->total_bytes) {
3717                         spin_unlock(&space_info->lock);
3718                         break;
3719                 }
3720                 spin_unlock(&space_info->lock);
3721
3722                 loops++;
3723                 if (wait_ordered && !trans) {
3724                         btrfs_wait_ordered_extents(root, 0, 0);
3725                 } else {
3726                         time_left = schedule_timeout_killable(1);
3727                         if (time_left)
3728                                 break;
3729                 }
3730                 smp_mb();
3731                 delalloc_bytes = root->fs_info->delalloc_bytes;
3732         }
3733 }
3734
3735 /**
3736  * maybe_commit_transaction - possibly commit the transaction if its ok to
3737  * @root - the root we're allocating for
3738  * @bytes - the number of bytes we want to reserve
3739  * @force - force the commit
3740  *
3741  * This will check to make sure that committing the transaction will actually
3742  * get us somewhere and then commit the transaction if it does.  Otherwise it
3743  * will return -ENOSPC.
3744  */
3745 static int may_commit_transaction(struct btrfs_root *root,
3746                                   struct btrfs_space_info *space_info,
3747                                   u64 bytes, int force)
3748 {
3749         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3750         struct btrfs_trans_handle *trans;
3751
3752         trans = (struct btrfs_trans_handle *)current->journal_info;
3753         if (trans)
3754                 return -EAGAIN;
3755
3756         if (force)
3757                 goto commit;
3758
3759         /* See if there is enough pinned space to make this reservation */
3760         spin_lock(&space_info->lock);
3761         if (space_info->bytes_pinned >= bytes) {
3762                 spin_unlock(&space_info->lock);
3763                 goto commit;
3764         }
3765         spin_unlock(&space_info->lock);
3766
3767         /*
3768          * See if there is some space in the delayed insertion reservation for
3769          * this reservation.
3770          */
3771         if (space_info != delayed_rsv->space_info)
3772                 return -ENOSPC;
3773
3774         spin_lock(&space_info->lock);
3775         spin_lock(&delayed_rsv->lock);
3776         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3777                 spin_unlock(&delayed_rsv->lock);
3778                 spin_unlock(&space_info->lock);
3779                 return -ENOSPC;
3780         }
3781         spin_unlock(&delayed_rsv->lock);
3782         spin_unlock(&space_info->lock);
3783
3784 commit:
3785         trans = btrfs_join_transaction(root);
3786         if (IS_ERR(trans))
3787                 return -ENOSPC;
3788
3789         return btrfs_commit_transaction(trans, root);
3790 }
3791
3792 enum flush_state {
3793         FLUSH_DELALLOC          =       1,
3794         FLUSH_DELALLOC_WAIT     =       2,
3795         FLUSH_DELAYED_ITEMS_NR  =       3,
3796         FLUSH_DELAYED_ITEMS     =       4,
3797         COMMIT_TRANS            =       5,
3798 };
3799
3800 static int flush_space(struct btrfs_root *root,
3801                        struct btrfs_space_info *space_info, u64 num_bytes,
3802                        u64 orig_bytes, int state)
3803 {
3804         struct btrfs_trans_handle *trans;
3805         int nr;
3806         int ret = 0;
3807
3808         switch (state) {
3809         case FLUSH_DELALLOC:
3810         case FLUSH_DELALLOC_WAIT:
3811                 shrink_delalloc(root, num_bytes, orig_bytes,
3812                                 state == FLUSH_DELALLOC_WAIT);
3813                 break;
3814         case FLUSH_DELAYED_ITEMS_NR:
3815         case FLUSH_DELAYED_ITEMS:
3816                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3817                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3818
3819                         nr = (int)div64_u64(num_bytes, bytes);
3820                         if (!nr)
3821                                 nr = 1;
3822                         nr *= 2;
3823                 } else {
3824                         nr = -1;
3825                 }
3826                 trans = btrfs_join_transaction(root);
3827                 if (IS_ERR(trans)) {
3828                         ret = PTR_ERR(trans);
3829                         break;
3830                 }
3831                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3832                 btrfs_end_transaction(trans, root);
3833                 break;
3834         case COMMIT_TRANS:
3835                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3836                 break;
3837         default:
3838                 ret = -ENOSPC;
3839                 break;
3840         }
3841
3842         return ret;
3843 }
3844 /**
3845  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3846  * @root - the root we're allocating for
3847  * @block_rsv - the block_rsv we're allocating for
3848  * @orig_bytes - the number of bytes we want
3849  * @flush - wether or not we can flush to make our reservation
3850  *
3851  * This will reserve orgi_bytes number of bytes from the space info associated
3852  * with the block_rsv.  If there is not enough space it will make an attempt to
3853  * flush out space to make room.  It will do this by flushing delalloc if
3854  * possible or committing the transaction.  If flush is 0 then no attempts to
3855  * regain reservations will be made and this will fail if there is not enough
3856  * space already.
3857  */
3858 static int reserve_metadata_bytes(struct btrfs_root *root,
3859                                   struct btrfs_block_rsv *block_rsv,
3860                                   u64 orig_bytes, int flush)
3861 {
3862         struct btrfs_space_info *space_info = block_rsv->space_info;
3863         u64 used;
3864         u64 num_bytes = orig_bytes;
3865         int flush_state = FLUSH_DELALLOC;
3866         int ret = 0;
3867         bool flushing = false;
3868         bool committed = false;
3869
3870 again:
3871         ret = 0;
3872         spin_lock(&space_info->lock);
3873         /*
3874          * We only want to wait if somebody other than us is flushing and we are
3875          * actually alloed to flush.
3876          */
3877         while (flush && !flushing && space_info->flush) {
3878                 spin_unlock(&space_info->lock);
3879                 /*
3880                  * If we have a trans handle we can't wait because the flusher
3881                  * may have to commit the transaction, which would mean we would
3882                  * deadlock since we are waiting for the flusher to finish, but
3883                  * hold the current transaction open.
3884                  */
3885                 if (current->journal_info)
3886                         return -EAGAIN;
3887                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3888                 /* Must have been killed, return */
3889                 if (ret)
3890                         return -EINTR;
3891
3892                 spin_lock(&space_info->lock);
3893         }
3894
3895         ret = -ENOSPC;
3896         used = space_info->bytes_used + space_info->bytes_reserved +
3897                 space_info->bytes_pinned + space_info->bytes_readonly +
3898                 space_info->bytes_may_use;
3899
3900         /*
3901          * The idea here is that we've not already over-reserved the block group
3902          * then we can go ahead and save our reservation first and then start
3903          * flushing if we need to.  Otherwise if we've already overcommitted
3904          * lets start flushing stuff first and then come back and try to make
3905          * our reservation.
3906          */
3907         if (used <= space_info->total_bytes) {
3908                 if (used + orig_bytes <= space_info->total_bytes) {
3909                         space_info->bytes_may_use += orig_bytes;
3910                         trace_btrfs_space_reservation(root->fs_info,
3911                                 "space_info", space_info->flags, orig_bytes, 1);
3912                         ret = 0;
3913                 } else {
3914                         /*
3915                          * Ok set num_bytes to orig_bytes since we aren't
3916                          * overocmmitted, this way we only try and reclaim what
3917                          * we need.
3918                          */
3919                         num_bytes = orig_bytes;
3920                 }
3921         } else {
3922                 /*
3923                  * Ok we're over committed, set num_bytes to the overcommitted
3924                  * amount plus the amount of bytes that we need for this
3925                  * reservation.
3926                  */
3927                 num_bytes = used - space_info->total_bytes +
3928                         (orig_bytes * 2);
3929         }
3930
3931         if (ret) {
3932                 u64 profile = btrfs_get_alloc_profile(root, 0);
3933                 u64 avail;
3934
3935                 /*
3936                  * If we have a lot of space that's pinned, don't bother doing
3937                  * the overcommit dance yet and just commit the transaction.
3938                  */
3939                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3940                 do_div(avail, 10);
3941                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3942                         space_info->flush = 1;
3943                         flushing = true;
3944                         spin_unlock(&space_info->lock);
3945                         ret = may_commit_transaction(root, space_info,
3946                                                      orig_bytes, 1);
3947                         if (ret)
3948                                 goto out;
3949                         committed = true;
3950                         goto again;
3951                 }
3952
3953                 spin_lock(&root->fs_info->free_chunk_lock);
3954                 avail = root->fs_info->free_chunk_space;
3955
3956                 /*
3957                  * If we have dup, raid1 or raid10 then only half of the free
3958                  * space is actually useable.
3959                  */
3960                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3961                                BTRFS_BLOCK_GROUP_RAID1 |
3962                                BTRFS_BLOCK_GROUP_RAID10))
3963                         avail >>= 1;
3964
3965                 /*
3966                  * If we aren't flushing don't let us overcommit too much, say
3967                  * 1/8th of the space.  If we can flush, let it overcommit up to
3968                  * 1/2 of the space.
3969                  */
3970                 if (flush)
3971                         avail >>= 3;
3972                 else
3973                         avail >>= 1;
3974                  spin_unlock(&root->fs_info->free_chunk_lock);
3975
3976                 if (used + num_bytes < space_info->total_bytes + avail) {
3977                         space_info->bytes_may_use += orig_bytes;
3978                         trace_btrfs_space_reservation(root->fs_info,
3979                                 "space_info", space_info->flags, orig_bytes, 1);
3980                         ret = 0;
3981                 }
3982         }
3983
3984         /*
3985          * Couldn't make our reservation, save our place so while we're trying
3986          * to reclaim space we can actually use it instead of somebody else
3987          * stealing it from us.
3988          */
3989         if (ret && flush) {
3990                 flushing = true;
3991                 space_info->flush = 1;
3992         }
3993
3994         spin_unlock(&space_info->lock);
3995
3996         if (!ret || !flush)
3997                 goto out;
3998
3999         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4000                           flush_state);
4001         flush_state++;
4002         if (!ret)
4003                 goto again;
4004         else if (flush_state <= COMMIT_TRANS)
4005                 goto again;
4006
4007 out:
4008         if (flushing) {
4009                 spin_lock(&space_info->lock);
4010                 space_info->flush = 0;
4011                 wake_up_all(&space_info->wait);
4012                 spin_unlock(&space_info->lock);
4013         }
4014         return ret;
4015 }
4016
4017 static struct btrfs_block_rsv *get_block_rsv(
4018                                         const struct btrfs_trans_handle *trans,
4019                                         const struct btrfs_root *root)
4020 {
4021         struct btrfs_block_rsv *block_rsv = NULL;
4022
4023         if (root->ref_cows)
4024                 block_rsv = trans->block_rsv;
4025
4026         if (root == root->fs_info->csum_root && trans->adding_csums)
4027                 block_rsv = trans->block_rsv;
4028
4029         if (!block_rsv)
4030                 block_rsv = root->block_rsv;
4031
4032         if (!block_rsv)
4033                 block_rsv = &root->fs_info->empty_block_rsv;
4034
4035         return block_rsv;
4036 }
4037
4038 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4039                                u64 num_bytes)
4040 {
4041         int ret = -ENOSPC;
4042         spin_lock(&block_rsv->lock);
4043         if (block_rsv->reserved >= num_bytes) {
4044                 block_rsv->reserved -= num_bytes;
4045                 if (block_rsv->reserved < block_rsv->size)
4046                         block_rsv->full = 0;
4047                 ret = 0;
4048         }
4049         spin_unlock(&block_rsv->lock);
4050         return ret;
4051 }
4052
4053 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4054                                 u64 num_bytes, int update_size)
4055 {
4056         spin_lock(&block_rsv->lock);
4057         block_rsv->reserved += num_bytes;
4058         if (update_size)
4059                 block_rsv->size += num_bytes;
4060         else if (block_rsv->reserved >= block_rsv->size)
4061                 block_rsv->full = 1;
4062         spin_unlock(&block_rsv->lock);
4063 }
4064
4065 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4066                                     struct btrfs_block_rsv *block_rsv,
4067                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4068 {
4069         struct btrfs_space_info *space_info = block_rsv->space_info;
4070
4071         spin_lock(&block_rsv->lock);
4072         if (num_bytes == (u64)-1)
4073                 num_bytes = block_rsv->size;
4074         block_rsv->size -= num_bytes;
4075         if (block_rsv->reserved >= block_rsv->size) {
4076                 num_bytes = block_rsv->reserved - block_rsv->size;
4077                 block_rsv->reserved = block_rsv->size;
4078                 block_rsv->full = 1;
4079         } else {
4080                 num_bytes = 0;
4081         }
4082         spin_unlock(&block_rsv->lock);
4083
4084         if (num_bytes > 0) {
4085                 if (dest) {
4086                         spin_lock(&dest->lock);
4087                         if (!dest->full) {
4088                                 u64 bytes_to_add;
4089
4090                                 bytes_to_add = dest->size - dest->reserved;
4091                                 bytes_to_add = min(num_bytes, bytes_to_add);
4092                                 dest->reserved += bytes_to_add;
4093                                 if (dest->reserved >= dest->size)
4094                                         dest->full = 1;
4095                                 num_bytes -= bytes_to_add;
4096                         }
4097                         spin_unlock(&dest->lock);
4098                 }
4099                 if (num_bytes) {
4100                         spin_lock(&space_info->lock);
4101                         space_info->bytes_may_use -= num_bytes;
4102                         trace_btrfs_space_reservation(fs_info, "space_info",
4103                                         space_info->flags, num_bytes, 0);
4104                         space_info->reservation_progress++;
4105                         spin_unlock(&space_info->lock);
4106                 }
4107         }
4108 }
4109
4110 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4111                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4112 {
4113         int ret;
4114
4115         ret = block_rsv_use_bytes(src, num_bytes);
4116         if (ret)
4117                 return ret;
4118
4119         block_rsv_add_bytes(dst, num_bytes, 1);
4120         return 0;
4121 }
4122
4123 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4124 {
4125         memset(rsv, 0, sizeof(*rsv));
4126         spin_lock_init(&rsv->lock);
4127 }
4128
4129 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4130 {
4131         struct btrfs_block_rsv *block_rsv;
4132         struct btrfs_fs_info *fs_info = root->fs_info;
4133
4134         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4135         if (!block_rsv)
4136                 return NULL;
4137
4138         btrfs_init_block_rsv(block_rsv);
4139         block_rsv->space_info = __find_space_info(fs_info,
4140                                                   BTRFS_BLOCK_GROUP_METADATA);
4141         return block_rsv;
4142 }
4143
4144 void btrfs_free_block_rsv(struct btrfs_root *root,
4145                           struct btrfs_block_rsv *rsv)
4146 {
4147         btrfs_block_rsv_release(root, rsv, (u64)-1);
4148         kfree(rsv);
4149 }
4150
4151 static inline int __block_rsv_add(struct btrfs_root *root,
4152                                   struct btrfs_block_rsv *block_rsv,
4153                                   u64 num_bytes, int flush)
4154 {
4155         int ret;
4156
4157         if (num_bytes == 0)
4158                 return 0;
4159
4160         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4161         if (!ret) {
4162                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4163                 return 0;
4164         }
4165
4166         return ret;
4167 }
4168
4169 int btrfs_block_rsv_add(struct btrfs_root *root,
4170                         struct btrfs_block_rsv *block_rsv,
4171                         u64 num_bytes)
4172 {
4173         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4174 }
4175
4176 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4177                                 struct btrfs_block_rsv *block_rsv,
4178                                 u64 num_bytes)
4179 {
4180         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4181 }
4182
4183 int btrfs_block_rsv_check(struct btrfs_root *root,
4184                           struct btrfs_block_rsv *block_rsv, int min_factor)
4185 {
4186         u64 num_bytes = 0;
4187         int ret = -ENOSPC;
4188
4189         if (!block_rsv)
4190                 return 0;
4191
4192         spin_lock(&block_rsv->lock);
4193         num_bytes = div_factor(block_rsv->size, min_factor);
4194         if (block_rsv->reserved >= num_bytes)
4195                 ret = 0;
4196         spin_unlock(&block_rsv->lock);
4197
4198         return ret;
4199 }
4200
4201 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4202                                            struct btrfs_block_rsv *block_rsv,
4203                                            u64 min_reserved, int flush)
4204 {
4205         u64 num_bytes = 0;
4206         int ret = -ENOSPC;
4207
4208         if (!block_rsv)
4209                 return 0;
4210
4211         spin_lock(&block_rsv->lock);
4212         num_bytes = min_reserved;
4213         if (block_rsv->reserved >= num_bytes)
4214                 ret = 0;
4215         else
4216                 num_bytes -= block_rsv->reserved;
4217         spin_unlock(&block_rsv->lock);
4218
4219         if (!ret)
4220                 return 0;
4221
4222         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4223         if (!ret) {
4224                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4225                 return 0;
4226         }
4227
4228         return ret;
4229 }
4230
4231 int btrfs_block_rsv_refill(struct btrfs_root *root,
4232                            struct btrfs_block_rsv *block_rsv,
4233                            u64 min_reserved)
4234 {
4235         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4236 }
4237
4238 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4239                                    struct btrfs_block_rsv *block_rsv,
4240                                    u64 min_reserved)
4241 {
4242         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4243 }
4244
4245 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4246                             struct btrfs_block_rsv *dst_rsv,
4247                             u64 num_bytes)
4248 {
4249         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4250 }
4251
4252 void btrfs_block_rsv_release(struct btrfs_root *root,
4253                              struct btrfs_block_rsv *block_rsv,
4254                              u64 num_bytes)
4255 {
4256         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4257         if (global_rsv->full || global_rsv == block_rsv ||
4258             block_rsv->space_info != global_rsv->space_info)
4259                 global_rsv = NULL;
4260         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4261                                 num_bytes);
4262 }
4263
4264 /*
4265  * helper to calculate size of global block reservation.
4266  * the desired value is sum of space used by extent tree,
4267  * checksum tree and root tree
4268  */
4269 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4270 {
4271         struct btrfs_space_info *sinfo;
4272         u64 num_bytes;
4273         u64 meta_used;
4274         u64 data_used;
4275         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4276
4277         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4278         spin_lock(&sinfo->lock);
4279         data_used = sinfo->bytes_used;
4280         spin_unlock(&sinfo->lock);
4281
4282         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4283         spin_lock(&sinfo->lock);
4284         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4285                 data_used = 0;
4286         meta_used = sinfo->bytes_used;
4287         spin_unlock(&sinfo->lock);
4288
4289         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4290                     csum_size * 2;
4291         num_bytes += div64_u64(data_used + meta_used, 50);
4292
4293         if (num_bytes * 3 > meta_used)
4294                 num_bytes = div64_u64(meta_used, 3);
4295
4296         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4297 }
4298
4299 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4300 {
4301         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4302         struct btrfs_space_info *sinfo = block_rsv->space_info;
4303         u64 num_bytes;
4304
4305         num_bytes = calc_global_metadata_size(fs_info);
4306
4307         spin_lock(&sinfo->lock);
4308         spin_lock(&block_rsv->lock);
4309
4310         block_rsv->size = num_bytes;
4311
4312         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4313                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4314                     sinfo->bytes_may_use;
4315
4316         if (sinfo->total_bytes > num_bytes) {
4317                 num_bytes = sinfo->total_bytes - num_bytes;
4318                 block_rsv->reserved += num_bytes;
4319                 sinfo->bytes_may_use += num_bytes;
4320                 trace_btrfs_space_reservation(fs_info, "space_info",
4321                                       sinfo->flags, num_bytes, 1);
4322         }
4323
4324         if (block_rsv->reserved >= block_rsv->size) {
4325                 num_bytes = block_rsv->reserved - block_rsv->size;
4326                 sinfo->bytes_may_use -= num_bytes;
4327                 trace_btrfs_space_reservation(fs_info, "space_info",
4328                                       sinfo->flags, num_bytes, 0);
4329                 sinfo->reservation_progress++;
4330                 block_rsv->reserved = block_rsv->size;
4331                 block_rsv->full = 1;
4332         }
4333
4334         spin_unlock(&block_rsv->lock);
4335         spin_unlock(&sinfo->lock);
4336 }
4337
4338 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4339 {
4340         struct btrfs_space_info *space_info;
4341
4342         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4343         fs_info->chunk_block_rsv.space_info = space_info;
4344
4345         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4346         fs_info->global_block_rsv.space_info = space_info;
4347         fs_info->delalloc_block_rsv.space_info = space_info;
4348         fs_info->trans_block_rsv.space_info = space_info;
4349         fs_info->empty_block_rsv.space_info = space_info;
4350         fs_info->delayed_block_rsv.space_info = space_info;
4351
4352         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4353         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4354         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4355         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4356         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4357
4358         update_global_block_rsv(fs_info);
4359 }
4360
4361 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4362 {
4363         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4364                                 (u64)-1);
4365         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4366         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4367         WARN_ON(fs_info->trans_block_rsv.size > 0);
4368         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4369         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4370         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4371         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4372         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4373 }
4374
4375 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4376                                   struct btrfs_root *root)
4377 {
4378         if (!trans->block_rsv)
4379                 return;
4380
4381         if (!trans->bytes_reserved)
4382                 return;
4383
4384         trace_btrfs_space_reservation(root->fs_info, "transaction",
4385                                       trans->transid, trans->bytes_reserved, 0);
4386         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4387         trans->bytes_reserved = 0;
4388 }
4389
4390 /* Can only return 0 or -ENOSPC */
4391 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4392                                   struct inode *inode)
4393 {
4394         struct btrfs_root *root = BTRFS_I(inode)->root;
4395         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4396         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4397
4398         /*
4399          * We need to hold space in order to delete our orphan item once we've
4400          * added it, so this takes the reservation so we can release it later
4401          * when we are truly done with the orphan item.
4402          */
4403         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4404         trace_btrfs_space_reservation(root->fs_info, "orphan",
4405                                       btrfs_ino(inode), num_bytes, 1);
4406         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4407 }
4408
4409 void btrfs_orphan_release_metadata(struct inode *inode)
4410 {
4411         struct btrfs_root *root = BTRFS_I(inode)->root;
4412         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4413         trace_btrfs_space_reservation(root->fs_info, "orphan",
4414                                       btrfs_ino(inode), num_bytes, 0);
4415         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4416 }
4417
4418 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4419                                 struct btrfs_pending_snapshot *pending)
4420 {
4421         struct btrfs_root *root = pending->root;
4422         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4423         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4424         /*
4425          * two for root back/forward refs, two for directory entries
4426          * and one for root of the snapshot.
4427          */
4428         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4429         dst_rsv->space_info = src_rsv->space_info;
4430         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4431 }
4432
4433 /**
4434  * drop_outstanding_extent - drop an outstanding extent
4435  * @inode: the inode we're dropping the extent for
4436  *
4437  * This is called when we are freeing up an outstanding extent, either called
4438  * after an error or after an extent is written.  This will return the number of
4439  * reserved extents that need to be freed.  This must be called with
4440  * BTRFS_I(inode)->lock held.
4441  */
4442 static unsigned drop_outstanding_extent(struct inode *inode)
4443 {
4444         unsigned drop_inode_space = 0;
4445         unsigned dropped_extents = 0;
4446
4447         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4448         BTRFS_I(inode)->outstanding_extents--;
4449
4450         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4451             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4452                                &BTRFS_I(inode)->runtime_flags))
4453                 drop_inode_space = 1;
4454
4455         /*
4456          * If we have more or the same amount of outsanding extents than we have
4457          * reserved then we need to leave the reserved extents count alone.
4458          */
4459         if (BTRFS_I(inode)->outstanding_extents >=
4460             BTRFS_I(inode)->reserved_extents)
4461                 return drop_inode_space;
4462
4463         dropped_extents = BTRFS_I(inode)->reserved_extents -
4464                 BTRFS_I(inode)->outstanding_extents;
4465         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4466         return dropped_extents + drop_inode_space;
4467 }
4468
4469 /**
4470  * calc_csum_metadata_size - return the amount of metada space that must be
4471  *      reserved/free'd for the given bytes.
4472  * @inode: the inode we're manipulating
4473  * @num_bytes: the number of bytes in question
4474  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4475  *
4476  * This adjusts the number of csum_bytes in the inode and then returns the
4477  * correct amount of metadata that must either be reserved or freed.  We
4478  * calculate how many checksums we can fit into one leaf and then divide the
4479  * number of bytes that will need to be checksumed by this value to figure out
4480  * how many checksums will be required.  If we are adding bytes then the number
4481  * may go up and we will return the number of additional bytes that must be
4482  * reserved.  If it is going down we will return the number of bytes that must
4483  * be freed.
4484  *
4485  * This must be called with BTRFS_I(inode)->lock held.
4486  */
4487 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4488                                    int reserve)
4489 {
4490         struct btrfs_root *root = BTRFS_I(inode)->root;
4491         u64 csum_size;
4492         int num_csums_per_leaf;
4493         int num_csums;
4494         int old_csums;
4495
4496         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4497             BTRFS_I(inode)->csum_bytes == 0)
4498                 return 0;
4499
4500         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4501         if (reserve)
4502                 BTRFS_I(inode)->csum_bytes += num_bytes;
4503         else
4504                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4505         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4506         num_csums_per_leaf = (int)div64_u64(csum_size,
4507                                             sizeof(struct btrfs_csum_item) +
4508                                             sizeof(struct btrfs_disk_key));
4509         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4510         num_csums = num_csums + num_csums_per_leaf - 1;
4511         num_csums = num_csums / num_csums_per_leaf;
4512
4513         old_csums = old_csums + num_csums_per_leaf - 1;
4514         old_csums = old_csums / num_csums_per_leaf;
4515
4516         /* No change, no need to reserve more */
4517         if (old_csums == num_csums)
4518                 return 0;
4519
4520         if (reserve)
4521                 return btrfs_calc_trans_metadata_size(root,
4522                                                       num_csums - old_csums);
4523
4524         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4525 }
4526
4527 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4528 {
4529         struct btrfs_root *root = BTRFS_I(inode)->root;
4530         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4531         u64 to_reserve = 0;
4532         u64 csum_bytes;
4533         unsigned nr_extents = 0;
4534         int extra_reserve = 0;
4535         int flush = 1;
4536         int ret;
4537
4538         /* Need to be holding the i_mutex here if we aren't free space cache */
4539         if (btrfs_is_free_space_inode(inode))
4540                 flush = 0;
4541
4542         if (flush && btrfs_transaction_in_commit(root->fs_info))
4543                 schedule_timeout(1);
4544
4545         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4546         num_bytes = ALIGN(num_bytes, root->sectorsize);
4547
4548         spin_lock(&BTRFS_I(inode)->lock);
4549         BTRFS_I(inode)->outstanding_extents++;
4550
4551         if (BTRFS_I(inode)->outstanding_extents >
4552             BTRFS_I(inode)->reserved_extents)
4553                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4554                         BTRFS_I(inode)->reserved_extents;
4555
4556         /*
4557          * Add an item to reserve for updating the inode when we complete the
4558          * delalloc io.
4559          */
4560         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4561                       &BTRFS_I(inode)->runtime_flags)) {
4562                 nr_extents++;
4563                 extra_reserve = 1;
4564         }
4565
4566         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4567         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4568         csum_bytes = BTRFS_I(inode)->csum_bytes;
4569         spin_unlock(&BTRFS_I(inode)->lock);
4570
4571         if (root->fs_info->quota_enabled) {
4572                 ret = btrfs_qgroup_reserve(root, num_bytes +
4573                                            nr_extents * root->leafsize);
4574                 if (ret) {
4575                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4576                         return ret;
4577                 }
4578         }
4579
4580         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4581         if (ret) {
4582                 u64 to_free = 0;
4583                 unsigned dropped;
4584
4585                 spin_lock(&BTRFS_I(inode)->lock);
4586                 dropped = drop_outstanding_extent(inode);
4587                 /*
4588                  * If the inodes csum_bytes is the same as the original
4589                  * csum_bytes then we know we haven't raced with any free()ers
4590                  * so we can just reduce our inodes csum bytes and carry on.
4591                  * Otherwise we have to do the normal free thing to account for
4592                  * the case that the free side didn't free up its reserve
4593                  * because of this outstanding reservation.
4594                  */
4595                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4596                         calc_csum_metadata_size(inode, num_bytes, 0);
4597                 else
4598                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4599                 spin_unlock(&BTRFS_I(inode)->lock);
4600                 if (dropped)
4601                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4602
4603                 if (to_free) {
4604                         btrfs_block_rsv_release(root, block_rsv, to_free);
4605                         trace_btrfs_space_reservation(root->fs_info,
4606                                                       "delalloc",
4607                                                       btrfs_ino(inode),
4608                                                       to_free, 0);
4609                 }
4610                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4611                 return ret;
4612         }
4613
4614         spin_lock(&BTRFS_I(inode)->lock);
4615         if (extra_reserve) {
4616                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4617                         &BTRFS_I(inode)->runtime_flags);
4618                 nr_extents--;
4619         }
4620         BTRFS_I(inode)->reserved_extents += nr_extents;
4621         spin_unlock(&BTRFS_I(inode)->lock);
4622         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4623
4624         if (to_reserve)
4625                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4626                                               btrfs_ino(inode), to_reserve, 1);
4627         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4628
4629         return 0;
4630 }
4631
4632 /**
4633  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4634  * @inode: the inode to release the reservation for
4635  * @num_bytes: the number of bytes we're releasing
4636  *
4637  * This will release the metadata reservation for an inode.  This can be called
4638  * once we complete IO for a given set of bytes to release their metadata
4639  * reservations.
4640  */
4641 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4642 {
4643         struct btrfs_root *root = BTRFS_I(inode)->root;
4644         u64 to_free = 0;
4645         unsigned dropped;
4646
4647         num_bytes = ALIGN(num_bytes, root->sectorsize);
4648         spin_lock(&BTRFS_I(inode)->lock);
4649         dropped = drop_outstanding_extent(inode);
4650
4651         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4652         spin_unlock(&BTRFS_I(inode)->lock);
4653         if (dropped > 0)
4654                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4655
4656         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4657                                       btrfs_ino(inode), to_free, 0);
4658         if (root->fs_info->quota_enabled) {
4659                 btrfs_qgroup_free(root, num_bytes +
4660                                         dropped * root->leafsize);
4661         }
4662
4663         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4664                                 to_free);
4665 }
4666
4667 /**
4668  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4669  * @inode: inode we're writing to
4670  * @num_bytes: the number of bytes we want to allocate
4671  *
4672  * This will do the following things
4673  *
4674  * o reserve space in the data space info for num_bytes
4675  * o reserve space in the metadata space info based on number of outstanding
4676  *   extents and how much csums will be needed
4677  * o add to the inodes ->delalloc_bytes
4678  * o add it to the fs_info's delalloc inodes list.
4679  *
4680  * This will return 0 for success and -ENOSPC if there is no space left.
4681  */
4682 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4683 {
4684         int ret;
4685
4686         ret = btrfs_check_data_free_space(inode, num_bytes);
4687         if (ret)
4688                 return ret;
4689
4690         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4691         if (ret) {
4692                 btrfs_free_reserved_data_space(inode, num_bytes);
4693                 return ret;
4694         }
4695
4696         return 0;
4697 }
4698
4699 /**
4700  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4701  * @inode: inode we're releasing space for
4702  * @num_bytes: the number of bytes we want to free up
4703  *
4704  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4705  * called in the case that we don't need the metadata AND data reservations
4706  * anymore.  So if there is an error or we insert an inline extent.
4707  *
4708  * This function will release the metadata space that was not used and will
4709  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4710  * list if there are no delalloc bytes left.
4711  */
4712 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4713 {
4714         btrfs_delalloc_release_metadata(inode, num_bytes);
4715         btrfs_free_reserved_data_space(inode, num_bytes);
4716 }
4717
4718 static int update_block_group(struct btrfs_trans_handle *trans,
4719                               struct btrfs_root *root,
4720                               u64 bytenr, u64 num_bytes, int alloc)
4721 {
4722         struct btrfs_block_group_cache *cache = NULL;
4723         struct btrfs_fs_info *info = root->fs_info;
4724         u64 total = num_bytes;
4725         u64 old_val;
4726         u64 byte_in_group;
4727         int factor;
4728
4729         /* block accounting for super block */
4730         spin_lock(&info->delalloc_lock);
4731         old_val = btrfs_super_bytes_used(info->super_copy);
4732         if (alloc)
4733                 old_val += num_bytes;
4734         else
4735                 old_val -= num_bytes;
4736         btrfs_set_super_bytes_used(info->super_copy, old_val);
4737         spin_unlock(&info->delalloc_lock);
4738
4739         while (total) {
4740                 cache = btrfs_lookup_block_group(info, bytenr);
4741                 if (!cache)
4742                         return -ENOENT;
4743                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4744                                     BTRFS_BLOCK_GROUP_RAID1 |
4745                                     BTRFS_BLOCK_GROUP_RAID10))
4746                         factor = 2;
4747                 else
4748                         factor = 1;
4749                 /*
4750                  * If this block group has free space cache written out, we
4751                  * need to make sure to load it if we are removing space.  This
4752                  * is because we need the unpinning stage to actually add the
4753                  * space back to the block group, otherwise we will leak space.
4754                  */
4755                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4756                         cache_block_group(cache, trans, NULL, 1);
4757
4758                 byte_in_group = bytenr - cache->key.objectid;
4759                 WARN_ON(byte_in_group > cache->key.offset);
4760
4761                 spin_lock(&cache->space_info->lock);
4762                 spin_lock(&cache->lock);
4763
4764                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4765                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4766                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4767
4768                 cache->dirty = 1;
4769                 old_val = btrfs_block_group_used(&cache->item);
4770                 num_bytes = min(total, cache->key.offset - byte_in_group);
4771                 if (alloc) {
4772                         old_val += num_bytes;
4773                         btrfs_set_block_group_used(&cache->item, old_val);
4774                         cache->reserved -= num_bytes;
4775                         cache->space_info->bytes_reserved -= num_bytes;
4776                         cache->space_info->bytes_used += num_bytes;
4777                         cache->space_info->disk_used += num_bytes * factor;
4778                         spin_unlock(&cache->lock);
4779                         spin_unlock(&cache->space_info->lock);
4780                 } else {
4781                         old_val -= num_bytes;
4782                         btrfs_set_block_group_used(&cache->item, old_val);
4783                         cache->pinned += num_bytes;
4784                         cache->space_info->bytes_pinned += num_bytes;
4785                         cache->space_info->bytes_used -= num_bytes;
4786                         cache->space_info->disk_used -= num_bytes * factor;
4787                         spin_unlock(&cache->lock);
4788                         spin_unlock(&cache->space_info->lock);
4789
4790                         set_extent_dirty(info->pinned_extents,
4791                                          bytenr, bytenr + num_bytes - 1,
4792                                          GFP_NOFS | __GFP_NOFAIL);
4793                 }
4794                 btrfs_put_block_group(cache);
4795                 total -= num_bytes;
4796                 bytenr += num_bytes;
4797         }
4798         return 0;
4799 }
4800
4801 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4802 {
4803         struct btrfs_block_group_cache *cache;
4804         u64 bytenr;
4805
4806         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4807         if (!cache)
4808                 return 0;
4809
4810         bytenr = cache->key.objectid;
4811         btrfs_put_block_group(cache);
4812
4813         return bytenr;
4814 }
4815
4816 static int pin_down_extent(struct btrfs_root *root,
4817                            struct btrfs_block_group_cache *cache,
4818                            u64 bytenr, u64 num_bytes, int reserved)
4819 {
4820         spin_lock(&cache->space_info->lock);
4821         spin_lock(&cache->lock);
4822         cache->pinned += num_bytes;
4823         cache->space_info->bytes_pinned += num_bytes;
4824         if (reserved) {
4825                 cache->reserved -= num_bytes;
4826                 cache->space_info->bytes_reserved -= num_bytes;
4827         }
4828         spin_unlock(&cache->lock);
4829         spin_unlock(&cache->space_info->lock);
4830
4831         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4832                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4833         return 0;
4834 }
4835
4836 /*
4837  * this function must be called within transaction
4838  */
4839 int btrfs_pin_extent(struct btrfs_root *root,
4840                      u64 bytenr, u64 num_bytes, int reserved)
4841 {
4842         struct btrfs_block_group_cache *cache;
4843
4844         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4845         BUG_ON(!cache); /* Logic error */
4846
4847         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4848
4849         btrfs_put_block_group(cache);
4850         return 0;
4851 }
4852
4853 /*
4854  * this function must be called within transaction
4855  */
4856 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4857                                     struct btrfs_root *root,
4858                                     u64 bytenr, u64 num_bytes)
4859 {
4860         struct btrfs_block_group_cache *cache;
4861
4862         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4863         BUG_ON(!cache); /* Logic error */
4864
4865         /*
4866          * pull in the free space cache (if any) so that our pin
4867          * removes the free space from the cache.  We have load_only set
4868          * to one because the slow code to read in the free extents does check
4869          * the pinned extents.
4870          */
4871         cache_block_group(cache, trans, root, 1);
4872
4873         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4874
4875         /* remove us from the free space cache (if we're there at all) */
4876         btrfs_remove_free_space(cache, bytenr, num_bytes);
4877         btrfs_put_block_group(cache);
4878         return 0;
4879 }
4880
4881 /**
4882  * btrfs_update_reserved_bytes - update the block_group and space info counters
4883  * @cache:      The cache we are manipulating
4884  * @num_bytes:  The number of bytes in question
4885  * @reserve:    One of the reservation enums
4886  *
4887  * This is called by the allocator when it reserves space, or by somebody who is
4888  * freeing space that was never actually used on disk.  For example if you
4889  * reserve some space for a new leaf in transaction A and before transaction A
4890  * commits you free that leaf, you call this with reserve set to 0 in order to
4891  * clear the reservation.
4892  *
4893  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4894  * ENOSPC accounting.  For data we handle the reservation through clearing the
4895  * delalloc bits in the io_tree.  We have to do this since we could end up
4896  * allocating less disk space for the amount of data we have reserved in the
4897  * case of compression.
4898  *
4899  * If this is a reservation and the block group has become read only we cannot
4900  * make the reservation and return -EAGAIN, otherwise this function always
4901  * succeeds.
4902  */
4903 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4904                                        u64 num_bytes, int reserve)
4905 {
4906         struct btrfs_space_info *space_info = cache->space_info;
4907         int ret = 0;
4908
4909         spin_lock(&space_info->lock);
4910         spin_lock(&cache->lock);
4911         if (reserve != RESERVE_FREE) {
4912                 if (cache->ro) {
4913                         ret = -EAGAIN;
4914                 } else {
4915                         cache->reserved += num_bytes;
4916                         space_info->bytes_reserved += num_bytes;
4917                         if (reserve == RESERVE_ALLOC) {
4918                                 trace_btrfs_space_reservation(cache->fs_info,
4919                                                 "space_info", space_info->flags,
4920                                                 num_bytes, 0);
4921                                 space_info->bytes_may_use -= num_bytes;
4922                         }
4923                 }
4924         } else {
4925                 if (cache->ro)
4926                         space_info->bytes_readonly += num_bytes;
4927                 cache->reserved -= num_bytes;
4928                 space_info->bytes_reserved -= num_bytes;
4929                 space_info->reservation_progress++;
4930         }
4931         spin_unlock(&cache->lock);
4932         spin_unlock(&space_info->lock);
4933         return ret;
4934 }
4935
4936 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4937                                 struct btrfs_root *root)
4938 {
4939         struct btrfs_fs_info *fs_info = root->fs_info;
4940         struct btrfs_caching_control *next;
4941         struct btrfs_caching_control *caching_ctl;
4942         struct btrfs_block_group_cache *cache;
4943
4944         down_write(&fs_info->extent_commit_sem);
4945
4946         list_for_each_entry_safe(caching_ctl, next,
4947                                  &fs_info->caching_block_groups, list) {
4948                 cache = caching_ctl->block_group;
4949                 if (block_group_cache_done(cache)) {
4950                         cache->last_byte_to_unpin = (u64)-1;
4951                         list_del_init(&caching_ctl->list);
4952                         put_caching_control(caching_ctl);
4953                 } else {
4954                         cache->last_byte_to_unpin = caching_ctl->progress;
4955                 }
4956         }
4957
4958         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4959                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4960         else
4961                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4962
4963         up_write(&fs_info->extent_commit_sem);
4964
4965         update_global_block_rsv(fs_info);
4966 }
4967
4968 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4969 {
4970         struct btrfs_fs_info *fs_info = root->fs_info;
4971         struct btrfs_block_group_cache *cache = NULL;
4972         u64 len;
4973
4974         while (start <= end) {
4975                 if (!cache ||
4976                     start >= cache->key.objectid + cache->key.offset) {
4977                         if (cache)
4978                                 btrfs_put_block_group(cache);
4979                         cache = btrfs_lookup_block_group(fs_info, start);
4980                         BUG_ON(!cache); /* Logic error */
4981                 }
4982
4983                 len = cache->key.objectid + cache->key.offset - start;
4984                 len = min(len, end + 1 - start);
4985
4986                 if (start < cache->last_byte_to_unpin) {
4987                         len = min(len, cache->last_byte_to_unpin - start);
4988                         btrfs_add_free_space(cache, start, len);
4989                 }
4990
4991                 start += len;
4992
4993                 spin_lock(&cache->space_info->lock);
4994                 spin_lock(&cache->lock);
4995                 cache->pinned -= len;
4996                 cache->space_info->bytes_pinned -= len;
4997                 if (cache->ro)
4998                         cache->space_info->bytes_readonly += len;
4999                 spin_unlock(&cache->lock);
5000                 spin_unlock(&cache->space_info->lock);
5001         }
5002
5003         if (cache)
5004                 btrfs_put_block_group(cache);
5005         return 0;
5006 }
5007
5008 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5009                                struct btrfs_root *root)
5010 {
5011         struct btrfs_fs_info *fs_info = root->fs_info;
5012         struct extent_io_tree *unpin;
5013         u64 start;
5014         u64 end;
5015         int ret;
5016
5017         if (trans->aborted)
5018                 return 0;
5019
5020         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5021                 unpin = &fs_info->freed_extents[1];
5022         else
5023                 unpin = &fs_info->freed_extents[0];
5024
5025         while (1) {
5026                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5027                                             EXTENT_DIRTY);
5028                 if (ret)
5029                         break;
5030
5031                 if (btrfs_test_opt(root, DISCARD))
5032                         ret = btrfs_discard_extent(root, start,
5033                                                    end + 1 - start, NULL);
5034
5035                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5036                 unpin_extent_range(root, start, end);
5037                 cond_resched();
5038         }
5039
5040         return 0;
5041 }
5042
5043 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5044                                 struct btrfs_root *root,
5045                                 u64 bytenr, u64 num_bytes, u64 parent,
5046                                 u64 root_objectid, u64 owner_objectid,
5047                                 u64 owner_offset, int refs_to_drop,
5048                                 struct btrfs_delayed_extent_op *extent_op)
5049 {
5050         struct btrfs_key key;
5051         struct btrfs_path *path;
5052         struct btrfs_fs_info *info = root->fs_info;
5053         struct btrfs_root *extent_root = info->extent_root;
5054         struct extent_buffer *leaf;
5055         struct btrfs_extent_item *ei;
5056         struct btrfs_extent_inline_ref *iref;
5057         int ret;
5058         int is_data;
5059         int extent_slot = 0;
5060         int found_extent = 0;
5061         int num_to_del = 1;
5062         u32 item_size;
5063         u64 refs;
5064
5065         path = btrfs_alloc_path();
5066         if (!path)
5067                 return -ENOMEM;
5068
5069         path->reada = 1;
5070         path->leave_spinning = 1;
5071
5072         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5073         BUG_ON(!is_data && refs_to_drop != 1);
5074
5075         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5076                                     bytenr, num_bytes, parent,
5077                                     root_objectid, owner_objectid,
5078                                     owner_offset);
5079         if (ret == 0) {
5080                 extent_slot = path->slots[0];
5081                 while (extent_slot >= 0) {
5082                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5083                                               extent_slot);
5084                         if (key.objectid != bytenr)
5085                                 break;
5086                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5087                             key.offset == num_bytes) {
5088                                 found_extent = 1;
5089                                 break;
5090                         }
5091                         if (path->slots[0] - extent_slot > 5)
5092                                 break;
5093                         extent_slot--;
5094                 }
5095 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5096                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5097                 if (found_extent && item_size < sizeof(*ei))
5098                         found_extent = 0;
5099 #endif
5100                 if (!found_extent) {
5101                         BUG_ON(iref);
5102                         ret = remove_extent_backref(trans, extent_root, path,
5103                                                     NULL, refs_to_drop,
5104                                                     is_data);
5105                         if (ret)
5106                                 goto abort;
5107                         btrfs_release_path(path);
5108                         path->leave_spinning = 1;
5109
5110                         key.objectid = bytenr;
5111                         key.type = BTRFS_EXTENT_ITEM_KEY;
5112                         key.offset = num_bytes;
5113
5114                         ret = btrfs_search_slot(trans, extent_root,
5115                                                 &key, path, -1, 1);
5116                         if (ret) {
5117                                 printk(KERN_ERR "umm, got %d back from search"
5118                                        ", was looking for %llu\n", ret,
5119                                        (unsigned long long)bytenr);
5120                                 if (ret > 0)
5121                                         btrfs_print_leaf(extent_root,
5122                                                          path->nodes[0]);
5123                         }
5124                         if (ret < 0)
5125                                 goto abort;
5126                         extent_slot = path->slots[0];
5127                 }
5128         } else if (ret == -ENOENT) {
5129                 btrfs_print_leaf(extent_root, path->nodes[0]);
5130                 WARN_ON(1);
5131                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5132                        "parent %llu root %llu  owner %llu offset %llu\n",
5133                        (unsigned long long)bytenr,
5134                        (unsigned long long)parent,
5135                        (unsigned long long)root_objectid,
5136                        (unsigned long long)owner_objectid,
5137                        (unsigned long long)owner_offset);
5138         } else {
5139                 goto abort;
5140         }
5141
5142         leaf = path->nodes[0];
5143         item_size = btrfs_item_size_nr(leaf, extent_slot);
5144 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5145         if (item_size < sizeof(*ei)) {
5146                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5147                 ret = convert_extent_item_v0(trans, extent_root, path,
5148                                              owner_objectid, 0);
5149                 if (ret < 0)
5150                         goto abort;
5151
5152                 btrfs_release_path(path);
5153                 path->leave_spinning = 1;
5154
5155                 key.objectid = bytenr;
5156                 key.type = BTRFS_EXTENT_ITEM_KEY;
5157                 key.offset = num_bytes;
5158
5159                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5160                                         -1, 1);
5161                 if (ret) {
5162                         printk(KERN_ERR "umm, got %d back from search"
5163                                ", was looking for %llu\n", ret,
5164                                (unsigned long long)bytenr);
5165                         btrfs_print_leaf(extent_root, path->nodes[0]);
5166                 }
5167                 if (ret < 0)
5168                         goto abort;
5169                 extent_slot = path->slots[0];
5170                 leaf = path->nodes[0];
5171                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5172         }
5173 #endif
5174         BUG_ON(item_size < sizeof(*ei));
5175         ei = btrfs_item_ptr(leaf, extent_slot,
5176                             struct btrfs_extent_item);
5177         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5178                 struct btrfs_tree_block_info *bi;
5179                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5180                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5181                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5182         }
5183
5184         refs = btrfs_extent_refs(leaf, ei);
5185         BUG_ON(refs < refs_to_drop);
5186         refs -= refs_to_drop;
5187
5188         if (refs > 0) {
5189                 if (extent_op)
5190                         __run_delayed_extent_op(extent_op, leaf, ei);
5191                 /*
5192                  * In the case of inline back ref, reference count will
5193                  * be updated by remove_extent_backref
5194                  */
5195                 if (iref) {
5196                         BUG_ON(!found_extent);
5197                 } else {
5198                         btrfs_set_extent_refs(leaf, ei, refs);
5199                         btrfs_mark_buffer_dirty(leaf);
5200                 }
5201                 if (found_extent) {
5202                         ret = remove_extent_backref(trans, extent_root, path,
5203                                                     iref, refs_to_drop,
5204                                                     is_data);
5205                         if (ret)
5206                                 goto abort;
5207                 }
5208         } else {
5209                 if (found_extent) {
5210                         BUG_ON(is_data && refs_to_drop !=
5211                                extent_data_ref_count(root, path, iref));
5212                         if (iref) {
5213                                 BUG_ON(path->slots[0] != extent_slot);
5214                         } else {
5215                                 BUG_ON(path->slots[0] != extent_slot + 1);
5216                                 path->slots[0] = extent_slot;
5217                                 num_to_del = 2;
5218                         }
5219                 }
5220
5221                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5222                                       num_to_del);
5223                 if (ret)
5224                         goto abort;
5225                 btrfs_release_path(path);
5226
5227                 if (is_data) {
5228                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5229                         if (ret)
5230                                 goto abort;
5231                 }
5232
5233                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5234                 if (ret)
5235                         goto abort;
5236         }
5237 out:
5238         btrfs_free_path(path);
5239         return ret;
5240
5241 abort:
5242         btrfs_abort_transaction(trans, extent_root, ret);
5243         goto out;
5244 }
5245
5246 /*
5247  * when we free an block, it is possible (and likely) that we free the last
5248  * delayed ref for that extent as well.  This searches the delayed ref tree for
5249  * a given extent, and if there are no other delayed refs to be processed, it
5250  * removes it from the tree.
5251  */
5252 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5253                                       struct btrfs_root *root, u64 bytenr)
5254 {
5255         struct btrfs_delayed_ref_head *head;
5256         struct btrfs_delayed_ref_root *delayed_refs;
5257         struct btrfs_delayed_ref_node *ref;
5258         struct rb_node *node;
5259         int ret = 0;
5260
5261         delayed_refs = &trans->transaction->delayed_refs;
5262         spin_lock(&delayed_refs->lock);
5263         head = btrfs_find_delayed_ref_head(trans, bytenr);
5264         if (!head)
5265                 goto out;
5266
5267         node = rb_prev(&head->node.rb_node);
5268         if (!node)
5269                 goto out;
5270
5271         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5272
5273         /* there are still entries for this ref, we can't drop it */
5274         if (ref->bytenr == bytenr)
5275                 goto out;
5276
5277         if (head->extent_op) {
5278                 if (!head->must_insert_reserved)
5279                         goto out;
5280                 kfree(head->extent_op);
5281                 head->extent_op = NULL;
5282         }
5283
5284         /*
5285          * waiting for the lock here would deadlock.  If someone else has it
5286          * locked they are already in the process of dropping it anyway
5287          */
5288         if (!mutex_trylock(&head->mutex))
5289                 goto out;
5290
5291         /*
5292          * at this point we have a head with no other entries.  Go
5293          * ahead and process it.
5294          */
5295         head->node.in_tree = 0;
5296         rb_erase(&head->node.rb_node, &delayed_refs->root);
5297
5298         delayed_refs->num_entries--;
5299         smp_mb();
5300         if (waitqueue_active(&root->fs_info->tree_mod_seq_wait))
5301                 wake_up(&root->fs_info->tree_mod_seq_wait);
5302
5303         /*
5304          * we don't take a ref on the node because we're removing it from the
5305          * tree, so we just steal the ref the tree was holding.
5306          */
5307         delayed_refs->num_heads--;
5308         if (list_empty(&head->cluster))
5309                 delayed_refs->num_heads_ready--;
5310
5311         list_del_init(&head->cluster);
5312         spin_unlock(&delayed_refs->lock);
5313
5314         BUG_ON(head->extent_op);
5315         if (head->must_insert_reserved)
5316                 ret = 1;
5317
5318         mutex_unlock(&head->mutex);
5319         btrfs_put_delayed_ref(&head->node);
5320         return ret;
5321 out:
5322         spin_unlock(&delayed_refs->lock);
5323         return 0;
5324 }
5325
5326 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5327                            struct btrfs_root *root,
5328                            struct extent_buffer *buf,
5329                            u64 parent, int last_ref)
5330 {
5331         struct btrfs_block_group_cache *cache = NULL;
5332         int ret;
5333
5334         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5335                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5336                                         buf->start, buf->len,
5337                                         parent, root->root_key.objectid,
5338                                         btrfs_header_level(buf),
5339                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5340                 BUG_ON(ret); /* -ENOMEM */
5341         }
5342
5343         if (!last_ref)
5344                 return;
5345
5346         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5347
5348         if (btrfs_header_generation(buf) == trans->transid) {
5349                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5350                         ret = check_ref_cleanup(trans, root, buf->start);
5351                         if (!ret)
5352                                 goto out;
5353                 }
5354
5355                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5356                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5357                         goto out;
5358                 }
5359
5360                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5361
5362                 btrfs_add_free_space(cache, buf->start, buf->len);
5363                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5364         }
5365 out:
5366         /*
5367          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5368          * anymore.
5369          */
5370         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5371         btrfs_put_block_group(cache);
5372 }
5373
5374 /* Can return -ENOMEM */
5375 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5376                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5377                       u64 owner, u64 offset, int for_cow)
5378 {
5379         int ret;
5380         struct btrfs_fs_info *fs_info = root->fs_info;
5381
5382         /*
5383          * tree log blocks never actually go into the extent allocation
5384          * tree, just update pinning info and exit early.
5385          */
5386         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5387                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5388                 /* unlocks the pinned mutex */
5389                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5390                 ret = 0;
5391         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5392                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5393                                         num_bytes,
5394                                         parent, root_objectid, (int)owner,
5395                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5396         } else {
5397                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5398                                                 num_bytes,
5399                                                 parent, root_objectid, owner,
5400                                                 offset, BTRFS_DROP_DELAYED_REF,
5401                                                 NULL, for_cow);
5402         }
5403         return ret;
5404 }
5405
5406 static u64 stripe_align(struct btrfs_root *root, u64 val)
5407 {
5408         u64 mask = ((u64)root->stripesize - 1);
5409         u64 ret = (val + mask) & ~mask;
5410         return ret;
5411 }
5412
5413 /*
5414  * when we wait for progress in the block group caching, its because
5415  * our allocation attempt failed at least once.  So, we must sleep
5416  * and let some progress happen before we try again.
5417  *
5418  * This function will sleep at least once waiting for new free space to
5419  * show up, and then it will check the block group free space numbers
5420  * for our min num_bytes.  Another option is to have it go ahead
5421  * and look in the rbtree for a free extent of a given size, but this
5422  * is a good start.
5423  */
5424 static noinline int
5425 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5426                                 u64 num_bytes)
5427 {
5428         struct btrfs_caching_control *caching_ctl;
5429         DEFINE_WAIT(wait);
5430
5431         caching_ctl = get_caching_control(cache);
5432         if (!caching_ctl)
5433                 return 0;
5434
5435         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5436                    (cache->free_space_ctl->free_space >= num_bytes));
5437
5438         put_caching_control(caching_ctl);
5439         return 0;
5440 }
5441
5442 static noinline int
5443 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5444 {
5445         struct btrfs_caching_control *caching_ctl;
5446         DEFINE_WAIT(wait);
5447
5448         caching_ctl = get_caching_control(cache);
5449         if (!caching_ctl)
5450                 return 0;
5451
5452         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5453
5454         put_caching_control(caching_ctl);
5455         return 0;
5456 }
5457
5458 static int __get_block_group_index(u64 flags)
5459 {
5460         int index;
5461
5462         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5463                 index = 0;
5464         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5465                 index = 1;
5466         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5467                 index = 2;
5468         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5469                 index = 3;
5470         else
5471                 index = 4;
5472
5473         return index;
5474 }
5475
5476 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5477 {
5478         return __get_block_group_index(cache->flags);
5479 }
5480
5481 enum btrfs_loop_type {
5482         LOOP_CACHING_NOWAIT = 0,
5483         LOOP_CACHING_WAIT = 1,
5484         LOOP_ALLOC_CHUNK = 2,
5485         LOOP_NO_EMPTY_SIZE = 3,
5486 };
5487
5488 /*
5489  * walks the btree of allocated extents and find a hole of a given size.
5490  * The key ins is changed to record the hole:
5491  * ins->objectid == block start
5492  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5493  * ins->offset == number of blocks
5494  * Any available blocks before search_start are skipped.
5495  */
5496 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5497                                      struct btrfs_root *orig_root,
5498                                      u64 num_bytes, u64 empty_size,
5499                                      u64 hint_byte, struct btrfs_key *ins,
5500                                      u64 data)
5501 {
5502         int ret = 0;
5503         struct btrfs_root *root = orig_root->fs_info->extent_root;
5504         struct btrfs_free_cluster *last_ptr = NULL;
5505         struct btrfs_block_group_cache *block_group = NULL;
5506         struct btrfs_block_group_cache *used_block_group;
5507         u64 search_start = 0;
5508         int empty_cluster = 2 * 1024 * 1024;
5509         int allowed_chunk_alloc = 0;
5510         int done_chunk_alloc = 0;
5511         struct btrfs_space_info *space_info;
5512         int loop = 0;
5513         int index = 0;
5514         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5515                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5516         bool found_uncached_bg = false;
5517         bool failed_cluster_refill = false;
5518         bool failed_alloc = false;
5519         bool use_cluster = true;
5520         bool have_caching_bg = false;
5521
5522         WARN_ON(num_bytes < root->sectorsize);
5523         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5524         ins->objectid = 0;
5525         ins->offset = 0;
5526
5527         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5528
5529         space_info = __find_space_info(root->fs_info, data);
5530         if (!space_info) {
5531                 printk(KERN_ERR "No space info for %llu\n", data);
5532                 return -ENOSPC;
5533         }
5534
5535         /*
5536          * If the space info is for both data and metadata it means we have a
5537          * small filesystem and we can't use the clustering stuff.
5538          */
5539         if (btrfs_mixed_space_info(space_info))
5540                 use_cluster = false;
5541
5542         if (orig_root->ref_cows || empty_size)
5543                 allowed_chunk_alloc = 1;
5544
5545         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5546                 last_ptr = &root->fs_info->meta_alloc_cluster;
5547                 if (!btrfs_test_opt(root, SSD))
5548                         empty_cluster = 64 * 1024;
5549         }
5550
5551         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5552             btrfs_test_opt(root, SSD)) {
5553                 last_ptr = &root->fs_info->data_alloc_cluster;
5554         }
5555
5556         if (last_ptr) {
5557                 spin_lock(&last_ptr->lock);
5558                 if (last_ptr->block_group)
5559                         hint_byte = last_ptr->window_start;
5560                 spin_unlock(&last_ptr->lock);
5561         }
5562
5563         search_start = max(search_start, first_logical_byte(root, 0));
5564         search_start = max(search_start, hint_byte);
5565
5566         if (!last_ptr)
5567                 empty_cluster = 0;
5568
5569         if (search_start == hint_byte) {
5570                 block_group = btrfs_lookup_block_group(root->fs_info,
5571                                                        search_start);
5572                 used_block_group = block_group;
5573                 /*
5574                  * we don't want to use the block group if it doesn't match our
5575                  * allocation bits, or if its not cached.
5576                  *
5577                  * However if we are re-searching with an ideal block group
5578                  * picked out then we don't care that the block group is cached.
5579                  */
5580                 if (block_group && block_group_bits(block_group, data) &&
5581                     block_group->cached != BTRFS_CACHE_NO) {
5582                         down_read(&space_info->groups_sem);
5583                         if (list_empty(&block_group->list) ||
5584                             block_group->ro) {
5585                                 /*
5586                                  * someone is removing this block group,
5587                                  * we can't jump into the have_block_group
5588                                  * target because our list pointers are not
5589                                  * valid
5590                                  */
5591                                 btrfs_put_block_group(block_group);
5592                                 up_read(&space_info->groups_sem);
5593                         } else {
5594                                 index = get_block_group_index(block_group);
5595                                 goto have_block_group;
5596                         }
5597                 } else if (block_group) {
5598                         btrfs_put_block_group(block_group);
5599                 }
5600         }
5601 search:
5602         have_caching_bg = false;
5603         down_read(&space_info->groups_sem);
5604         list_for_each_entry(block_group, &space_info->block_groups[index],
5605                             list) {
5606                 u64 offset;
5607                 int cached;
5608
5609                 used_block_group = block_group;
5610                 btrfs_get_block_group(block_group);
5611                 search_start = block_group->key.objectid;
5612
5613                 /*
5614                  * this can happen if we end up cycling through all the
5615                  * raid types, but we want to make sure we only allocate
5616                  * for the proper type.
5617                  */
5618                 if (!block_group_bits(block_group, data)) {
5619                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5620                                 BTRFS_BLOCK_GROUP_RAID1 |
5621                                 BTRFS_BLOCK_GROUP_RAID10;
5622
5623                         /*
5624                          * if they asked for extra copies and this block group
5625                          * doesn't provide them, bail.  This does allow us to
5626                          * fill raid0 from raid1.
5627                          */
5628                         if ((data & extra) && !(block_group->flags & extra))
5629                                 goto loop;
5630                 }
5631
5632 have_block_group:
5633                 cached = block_group_cache_done(block_group);
5634                 if (unlikely(!cached)) {
5635                         found_uncached_bg = true;
5636                         ret = cache_block_group(block_group, trans,
5637                                                 orig_root, 0);
5638                         BUG_ON(ret < 0);
5639                         ret = 0;
5640                 }
5641
5642                 if (unlikely(block_group->ro))
5643                         goto loop;
5644
5645                 /*
5646                  * Ok we want to try and use the cluster allocator, so
5647                  * lets look there
5648                  */
5649                 if (last_ptr) {
5650                         /*
5651                          * the refill lock keeps out other
5652                          * people trying to start a new cluster
5653                          */
5654                         spin_lock(&last_ptr->refill_lock);
5655                         used_block_group = last_ptr->block_group;
5656                         if (used_block_group != block_group &&
5657                             (!used_block_group ||
5658                              used_block_group->ro ||
5659                              !block_group_bits(used_block_group, data))) {
5660                                 used_block_group = block_group;
5661                                 goto refill_cluster;
5662                         }
5663
5664                         if (used_block_group != block_group)
5665                                 btrfs_get_block_group(used_block_group);
5666
5667                         offset = btrfs_alloc_from_cluster(used_block_group,
5668                           last_ptr, num_bytes, used_block_group->key.objectid);
5669                         if (offset) {
5670                                 /* we have a block, we're done */
5671                                 spin_unlock(&last_ptr->refill_lock);
5672                                 trace_btrfs_reserve_extent_cluster(root,
5673                                         block_group, search_start, num_bytes);
5674                                 goto checks;
5675                         }
5676
5677                         WARN_ON(last_ptr->block_group != used_block_group);
5678                         if (used_block_group != block_group) {
5679                                 btrfs_put_block_group(used_block_group);
5680                                 used_block_group = block_group;
5681                         }
5682 refill_cluster:
5683                         BUG_ON(used_block_group != block_group);
5684                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5685                          * set up a new clusters, so lets just skip it
5686                          * and let the allocator find whatever block
5687                          * it can find.  If we reach this point, we
5688                          * will have tried the cluster allocator
5689                          * plenty of times and not have found
5690                          * anything, so we are likely way too
5691                          * fragmented for the clustering stuff to find
5692                          * anything.
5693                          *
5694                          * However, if the cluster is taken from the
5695                          * current block group, release the cluster
5696                          * first, so that we stand a better chance of
5697                          * succeeding in the unclustered
5698                          * allocation.  */
5699                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5700                             last_ptr->block_group != block_group) {
5701                                 spin_unlock(&last_ptr->refill_lock);
5702                                 goto unclustered_alloc;
5703                         }
5704
5705                         /*
5706                          * this cluster didn't work out, free it and
5707                          * start over
5708                          */
5709                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5710
5711                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5712                                 spin_unlock(&last_ptr->refill_lock);
5713                                 goto unclustered_alloc;
5714                         }
5715
5716                         /* allocate a cluster in this block group */
5717                         ret = btrfs_find_space_cluster(trans, root,
5718                                                block_group, last_ptr,
5719                                                search_start, num_bytes,
5720                                                empty_cluster + empty_size);
5721                         if (ret == 0) {
5722                                 /*
5723                                  * now pull our allocation out of this
5724                                  * cluster
5725                                  */
5726                                 offset = btrfs_alloc_from_cluster(block_group,
5727                                                   last_ptr, num_bytes,
5728                                                   search_start);
5729                                 if (offset) {
5730                                         /* we found one, proceed */
5731                                         spin_unlock(&last_ptr->refill_lock);
5732                                         trace_btrfs_reserve_extent_cluster(root,
5733                                                 block_group, search_start,
5734                                                 num_bytes);
5735                                         goto checks;
5736                                 }
5737                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5738                                    && !failed_cluster_refill) {
5739                                 spin_unlock(&last_ptr->refill_lock);
5740
5741                                 failed_cluster_refill = true;
5742                                 wait_block_group_cache_progress(block_group,
5743                                        num_bytes + empty_cluster + empty_size);
5744                                 goto have_block_group;
5745                         }
5746
5747                         /*
5748                          * at this point we either didn't find a cluster
5749                          * or we weren't able to allocate a block from our
5750                          * cluster.  Free the cluster we've been trying
5751                          * to use, and go to the next block group
5752                          */
5753                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5754                         spin_unlock(&last_ptr->refill_lock);
5755                         goto loop;
5756                 }
5757
5758 unclustered_alloc:
5759                 spin_lock(&block_group->free_space_ctl->tree_lock);
5760                 if (cached &&
5761                     block_group->free_space_ctl->free_space <
5762                     num_bytes + empty_cluster + empty_size) {
5763                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5764                         goto loop;
5765                 }
5766                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5767
5768                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5769                                                     num_bytes, empty_size);
5770                 /*
5771                  * If we didn't find a chunk, and we haven't failed on this
5772                  * block group before, and this block group is in the middle of
5773                  * caching and we are ok with waiting, then go ahead and wait
5774                  * for progress to be made, and set failed_alloc to true.
5775                  *
5776                  * If failed_alloc is true then we've already waited on this
5777                  * block group once and should move on to the next block group.
5778                  */
5779                 if (!offset && !failed_alloc && !cached &&
5780                     loop > LOOP_CACHING_NOWAIT) {
5781                         wait_block_group_cache_progress(block_group,
5782                                                 num_bytes + empty_size);
5783                         failed_alloc = true;
5784                         goto have_block_group;
5785                 } else if (!offset) {
5786                         if (!cached)
5787                                 have_caching_bg = true;
5788                         goto loop;
5789                 }
5790 checks:
5791                 search_start = stripe_align(root, offset);
5792
5793                 /* move on to the next group */
5794                 if (search_start + num_bytes >
5795                     used_block_group->key.objectid + used_block_group->key.offset) {
5796                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5797                         goto loop;
5798                 }
5799
5800                 if (offset < search_start)
5801                         btrfs_add_free_space(used_block_group, offset,
5802                                              search_start - offset);
5803                 BUG_ON(offset > search_start);
5804
5805                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5806                                                   alloc_type);
5807                 if (ret == -EAGAIN) {
5808                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5809                         goto loop;
5810                 }
5811
5812                 /* we are all good, lets return */
5813                 ins->objectid = search_start;
5814                 ins->offset = num_bytes;
5815
5816                 trace_btrfs_reserve_extent(orig_root, block_group,
5817                                            search_start, num_bytes);
5818                 if (offset < search_start)
5819                         btrfs_add_free_space(used_block_group, offset,
5820                                              search_start - offset);
5821                 BUG_ON(offset > search_start);
5822                 if (used_block_group != block_group)
5823                         btrfs_put_block_group(used_block_group);
5824                 btrfs_put_block_group(block_group);
5825                 break;
5826 loop:
5827                 failed_cluster_refill = false;
5828                 failed_alloc = false;
5829                 BUG_ON(index != get_block_group_index(block_group));
5830                 if (used_block_group != block_group)
5831                         btrfs_put_block_group(used_block_group);
5832                 btrfs_put_block_group(block_group);
5833         }
5834         up_read(&space_info->groups_sem);
5835
5836         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5837                 goto search;
5838
5839         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5840                 goto search;
5841
5842         /*
5843          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5844          *                      caching kthreads as we move along
5845          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5846          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5847          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5848          *                      again
5849          */
5850         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5851                 index = 0;
5852                 loop++;
5853                 if (loop == LOOP_ALLOC_CHUNK) {
5854                        if (allowed_chunk_alloc) {
5855                                 ret = do_chunk_alloc(trans, root, num_bytes +
5856                                                      2 * 1024 * 1024, data,
5857                                                      CHUNK_ALLOC_LIMITED);
5858                                 /*
5859                                  * Do not bail out on ENOSPC since we
5860                                  * can do more things.
5861                                  */
5862                                 if (ret < 0 && ret != -ENOSPC) {
5863                                         btrfs_abort_transaction(trans,
5864                                                                 root, ret);
5865                                         goto out;
5866                                 }
5867                                 allowed_chunk_alloc = 0;
5868                                 if (ret == 1)
5869                                         done_chunk_alloc = 1;
5870                         } else if (!done_chunk_alloc &&
5871                                    space_info->force_alloc ==
5872                                    CHUNK_ALLOC_NO_FORCE) {
5873                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5874                         }
5875
5876                        /*
5877                         * We didn't allocate a chunk, go ahead and drop the
5878                         * empty size and loop again.
5879                         */
5880                        if (!done_chunk_alloc)
5881                                loop = LOOP_NO_EMPTY_SIZE;
5882                 }
5883
5884                 if (loop == LOOP_NO_EMPTY_SIZE) {
5885                         empty_size = 0;
5886                         empty_cluster = 0;
5887                 }
5888
5889                 goto search;
5890         } else if (!ins->objectid) {
5891                 ret = -ENOSPC;
5892         } else if (ins->objectid) {
5893                 ret = 0;
5894         }
5895 out:
5896
5897         return ret;
5898 }
5899
5900 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5901                             int dump_block_groups)
5902 {
5903         struct btrfs_block_group_cache *cache;
5904         int index = 0;
5905
5906         spin_lock(&info->lock);
5907         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5908                (unsigned long long)info->flags,
5909                (unsigned long long)(info->total_bytes - info->bytes_used -
5910                                     info->bytes_pinned - info->bytes_reserved -
5911                                     info->bytes_readonly),
5912                (info->full) ? "" : "not ");
5913         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5914                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5915                (unsigned long long)info->total_bytes,
5916                (unsigned long long)info->bytes_used,
5917                (unsigned long long)info->bytes_pinned,
5918                (unsigned long long)info->bytes_reserved,
5919                (unsigned long long)info->bytes_may_use,
5920                (unsigned long long)info->bytes_readonly);
5921         spin_unlock(&info->lock);
5922
5923         if (!dump_block_groups)
5924                 return;
5925
5926         down_read(&info->groups_sem);
5927 again:
5928         list_for_each_entry(cache, &info->block_groups[index], list) {
5929                 spin_lock(&cache->lock);
5930                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5931                        (unsigned long long)cache->key.objectid,
5932                        (unsigned long long)cache->key.offset,
5933                        (unsigned long long)btrfs_block_group_used(&cache->item),
5934                        (unsigned long long)cache->pinned,
5935                        (unsigned long long)cache->reserved,
5936                        cache->ro ? "[readonly]" : "");
5937                 btrfs_dump_free_space(cache, bytes);
5938                 spin_unlock(&cache->lock);
5939         }
5940         if (++index < BTRFS_NR_RAID_TYPES)
5941                 goto again;
5942         up_read(&info->groups_sem);
5943 }
5944
5945 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5946                          struct btrfs_root *root,
5947                          u64 num_bytes, u64 min_alloc_size,
5948                          u64 empty_size, u64 hint_byte,
5949                          struct btrfs_key *ins, u64 data)
5950 {
5951         bool final_tried = false;
5952         int ret;
5953
5954         data = btrfs_get_alloc_profile(root, data);
5955 again:
5956         /*
5957          * the only place that sets empty_size is btrfs_realloc_node, which
5958          * is not called recursively on allocations
5959          */
5960         if (empty_size || root->ref_cows) {
5961                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5962                                      num_bytes + 2 * 1024 * 1024, data,
5963                                      CHUNK_ALLOC_NO_FORCE);
5964                 if (ret < 0 && ret != -ENOSPC) {
5965                         btrfs_abort_transaction(trans, root, ret);
5966                         return ret;
5967                 }
5968         }
5969
5970         WARN_ON(num_bytes < root->sectorsize);
5971         ret = find_free_extent(trans, root, num_bytes, empty_size,
5972                                hint_byte, ins, data);
5973
5974         if (ret == -ENOSPC) {
5975                 if (!final_tried) {
5976                         num_bytes = num_bytes >> 1;
5977                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5978                         num_bytes = max(num_bytes, min_alloc_size);
5979                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5980                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5981                         if (ret < 0 && ret != -ENOSPC) {
5982                                 btrfs_abort_transaction(trans, root, ret);
5983                                 return ret;
5984                         }
5985                         if (num_bytes == min_alloc_size)
5986                                 final_tried = true;
5987                         goto again;
5988                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5989                         struct btrfs_space_info *sinfo;
5990
5991                         sinfo = __find_space_info(root->fs_info, data);
5992                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5993                                "wanted %llu\n", (unsigned long long)data,
5994                                (unsigned long long)num_bytes);
5995                         if (sinfo)
5996                                 dump_space_info(sinfo, num_bytes, 1);
5997                 }
5998         }
5999
6000         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6001
6002         return ret;
6003 }
6004
6005 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6006                                         u64 start, u64 len, int pin)
6007 {
6008         struct btrfs_block_group_cache *cache;
6009         int ret = 0;
6010
6011         cache = btrfs_lookup_block_group(root->fs_info, start);
6012         if (!cache) {
6013                 printk(KERN_ERR "Unable to find block group for %llu\n",
6014                        (unsigned long long)start);
6015                 return -ENOSPC;
6016         }
6017
6018         if (btrfs_test_opt(root, DISCARD))
6019                 ret = btrfs_discard_extent(root, start, len, NULL);
6020
6021         if (pin)
6022                 pin_down_extent(root, cache, start, len, 1);
6023         else {
6024                 btrfs_add_free_space(cache, start, len);
6025                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6026         }
6027         btrfs_put_block_group(cache);
6028
6029         trace_btrfs_reserved_extent_free(root, start, len);
6030
6031         return ret;
6032 }
6033
6034 int btrfs_free_reserved_extent(struct btrfs_root *root,
6035                                         u64 start, u64 len)
6036 {
6037         return __btrfs_free_reserved_extent(root, start, len, 0);
6038 }
6039
6040 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6041                                        u64 start, u64 len)
6042 {
6043         return __btrfs_free_reserved_extent(root, start, len, 1);
6044 }
6045
6046 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6047                                       struct btrfs_root *root,
6048                                       u64 parent, u64 root_objectid,
6049                                       u64 flags, u64 owner, u64 offset,
6050                                       struct btrfs_key *ins, int ref_mod)
6051 {
6052         int ret;
6053         struct btrfs_fs_info *fs_info = root->fs_info;
6054         struct btrfs_extent_item *extent_item;
6055         struct btrfs_extent_inline_ref *iref;
6056         struct btrfs_path *path;
6057         struct extent_buffer *leaf;
6058         int type;
6059         u32 size;
6060
6061         if (parent > 0)
6062                 type = BTRFS_SHARED_DATA_REF_KEY;
6063         else
6064                 type = BTRFS_EXTENT_DATA_REF_KEY;
6065
6066         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6067
6068         path = btrfs_alloc_path();
6069         if (!path)
6070                 return -ENOMEM;
6071
6072         path->leave_spinning = 1;
6073         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6074                                       ins, size);
6075         if (ret) {
6076                 btrfs_free_path(path);
6077                 return ret;
6078         }
6079
6080         leaf = path->nodes[0];
6081         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6082                                      struct btrfs_extent_item);
6083         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6084         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6085         btrfs_set_extent_flags(leaf, extent_item,
6086                                flags | BTRFS_EXTENT_FLAG_DATA);
6087
6088         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6089         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6090         if (parent > 0) {
6091                 struct btrfs_shared_data_ref *ref;
6092                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6093                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6094                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6095         } else {
6096                 struct btrfs_extent_data_ref *ref;
6097                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6098                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6099                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6100                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6101                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6102         }
6103
6104         btrfs_mark_buffer_dirty(path->nodes[0]);
6105         btrfs_free_path(path);
6106
6107         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6108         if (ret) { /* -ENOENT, logic error */
6109                 printk(KERN_ERR "btrfs update block group failed for %llu "
6110                        "%llu\n", (unsigned long long)ins->objectid,
6111                        (unsigned long long)ins->offset);
6112                 BUG();
6113         }
6114         return ret;
6115 }
6116
6117 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6118                                      struct btrfs_root *root,
6119                                      u64 parent, u64 root_objectid,
6120                                      u64 flags, struct btrfs_disk_key *key,
6121                                      int level, struct btrfs_key *ins)
6122 {
6123         int ret;
6124         struct btrfs_fs_info *fs_info = root->fs_info;
6125         struct btrfs_extent_item *extent_item;
6126         struct btrfs_tree_block_info *block_info;
6127         struct btrfs_extent_inline_ref *iref;
6128         struct btrfs_path *path;
6129         struct extent_buffer *leaf;
6130         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6131
6132         path = btrfs_alloc_path();
6133         if (!path)
6134                 return -ENOMEM;
6135
6136         path->leave_spinning = 1;
6137         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6138                                       ins, size);
6139         if (ret) {
6140                 btrfs_free_path(path);
6141                 return ret;
6142         }
6143
6144         leaf = path->nodes[0];
6145         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6146                                      struct btrfs_extent_item);
6147         btrfs_set_extent_refs(leaf, extent_item, 1);
6148         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6149         btrfs_set_extent_flags(leaf, extent_item,
6150                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6151         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6152
6153         btrfs_set_tree_block_key(leaf, block_info, key);
6154         btrfs_set_tree_block_level(leaf, block_info, level);
6155
6156         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6157         if (parent > 0) {
6158                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6159                 btrfs_set_extent_inline_ref_type(leaf, iref,
6160                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6161                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6162         } else {
6163                 btrfs_set_extent_inline_ref_type(leaf, iref,
6164                                                  BTRFS_TREE_BLOCK_REF_KEY);
6165                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6166         }
6167
6168         btrfs_mark_buffer_dirty(leaf);
6169         btrfs_free_path(path);
6170
6171         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6172         if (ret) { /* -ENOENT, logic error */
6173                 printk(KERN_ERR "btrfs update block group failed for %llu "
6174                        "%llu\n", (unsigned long long)ins->objectid,
6175                        (unsigned long long)ins->offset);
6176                 BUG();
6177         }
6178         return ret;
6179 }
6180
6181 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6182                                      struct btrfs_root *root,
6183                                      u64 root_objectid, u64 owner,
6184                                      u64 offset, struct btrfs_key *ins)
6185 {
6186         int ret;
6187
6188         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6189
6190         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6191                                          ins->offset, 0,
6192                                          root_objectid, owner, offset,
6193                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6194         return ret;
6195 }
6196
6197 /*
6198  * this is used by the tree logging recovery code.  It records that
6199  * an extent has been allocated and makes sure to clear the free
6200  * space cache bits as well
6201  */
6202 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6203                                    struct btrfs_root *root,
6204                                    u64 root_objectid, u64 owner, u64 offset,
6205                                    struct btrfs_key *ins)
6206 {
6207         int ret;
6208         struct btrfs_block_group_cache *block_group;
6209         struct btrfs_caching_control *caching_ctl;
6210         u64 start = ins->objectid;
6211         u64 num_bytes = ins->offset;
6212
6213         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6214         cache_block_group(block_group, trans, NULL, 0);
6215         caching_ctl = get_caching_control(block_group);
6216
6217         if (!caching_ctl) {
6218                 BUG_ON(!block_group_cache_done(block_group));
6219                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6220                 BUG_ON(ret); /* -ENOMEM */
6221         } else {
6222                 mutex_lock(&caching_ctl->mutex);
6223
6224                 if (start >= caching_ctl->progress) {
6225                         ret = add_excluded_extent(root, start, num_bytes);
6226                         BUG_ON(ret); /* -ENOMEM */
6227                 } else if (start + num_bytes <= caching_ctl->progress) {
6228                         ret = btrfs_remove_free_space(block_group,
6229                                                       start, num_bytes);
6230                         BUG_ON(ret); /* -ENOMEM */
6231                 } else {
6232                         num_bytes = caching_ctl->progress - start;
6233                         ret = btrfs_remove_free_space(block_group,
6234                                                       start, num_bytes);
6235                         BUG_ON(ret); /* -ENOMEM */
6236
6237                         start = caching_ctl->progress;
6238                         num_bytes = ins->objectid + ins->offset -
6239                                     caching_ctl->progress;
6240                         ret = add_excluded_extent(root, start, num_bytes);
6241                         BUG_ON(ret); /* -ENOMEM */
6242                 }
6243
6244                 mutex_unlock(&caching_ctl->mutex);
6245                 put_caching_control(caching_ctl);
6246         }
6247
6248         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6249                                           RESERVE_ALLOC_NO_ACCOUNT);
6250         BUG_ON(ret); /* logic error */
6251         btrfs_put_block_group(block_group);
6252         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6253                                          0, owner, offset, ins, 1);
6254         return ret;
6255 }
6256
6257 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6258                                             struct btrfs_root *root,
6259                                             u64 bytenr, u32 blocksize,
6260                                             int level)
6261 {
6262         struct extent_buffer *buf;
6263
6264         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6265         if (!buf)
6266                 return ERR_PTR(-ENOMEM);
6267         btrfs_set_header_generation(buf, trans->transid);
6268         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6269         btrfs_tree_lock(buf);
6270         clean_tree_block(trans, root, buf);
6271         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6272
6273         btrfs_set_lock_blocking(buf);
6274         btrfs_set_buffer_uptodate(buf);
6275
6276         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6277                 /*
6278                  * we allow two log transactions at a time, use different
6279                  * EXENT bit to differentiate dirty pages.
6280                  */
6281                 if (root->log_transid % 2 == 0)
6282                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6283                                         buf->start + buf->len - 1, GFP_NOFS);
6284                 else
6285                         set_extent_new(&root->dirty_log_pages, buf->start,
6286                                         buf->start + buf->len - 1, GFP_NOFS);
6287         } else {
6288                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6289                          buf->start + buf->len - 1, GFP_NOFS);
6290         }
6291         trans->blocks_used++;
6292         /* this returns a buffer locked for blocking */
6293         return buf;
6294 }
6295
6296 static struct btrfs_block_rsv *
6297 use_block_rsv(struct btrfs_trans_handle *trans,
6298               struct btrfs_root *root, u32 blocksize)
6299 {
6300         struct btrfs_block_rsv *block_rsv;
6301         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6302         int ret;
6303
6304         block_rsv = get_block_rsv(trans, root);
6305
6306         if (block_rsv->size == 0) {
6307                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6308                 /*
6309                  * If we couldn't reserve metadata bytes try and use some from
6310                  * the global reserve.
6311                  */
6312                 if (ret && block_rsv != global_rsv) {
6313                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6314                         if (!ret)
6315                                 return global_rsv;
6316                         return ERR_PTR(ret);
6317                 } else if (ret) {
6318                         return ERR_PTR(ret);
6319                 }
6320                 return block_rsv;
6321         }
6322
6323         ret = block_rsv_use_bytes(block_rsv, blocksize);
6324         if (!ret)
6325                 return block_rsv;
6326         if (ret) {
6327                 static DEFINE_RATELIMIT_STATE(_rs,
6328                                 DEFAULT_RATELIMIT_INTERVAL,
6329                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6330                 if (__ratelimit(&_rs)) {
6331                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6332                         WARN_ON(1);
6333                 }
6334                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6335                 if (!ret) {
6336                         return block_rsv;
6337                 } else if (ret && block_rsv != global_rsv) {
6338                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6339                         if (!ret)
6340                                 return global_rsv;
6341                 }
6342         }
6343
6344         return ERR_PTR(-ENOSPC);
6345 }
6346
6347 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6348                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6349 {
6350         block_rsv_add_bytes(block_rsv, blocksize, 0);
6351         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6352 }
6353
6354 /*
6355  * finds a free extent and does all the dirty work required for allocation
6356  * returns the key for the extent through ins, and a tree buffer for
6357  * the first block of the extent through buf.
6358  *
6359  * returns the tree buffer or NULL.
6360  */
6361 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6362                                         struct btrfs_root *root, u32 blocksize,
6363                                         u64 parent, u64 root_objectid,
6364                                         struct btrfs_disk_key *key, int level,
6365                                         u64 hint, u64 empty_size)
6366 {
6367         struct btrfs_key ins;
6368         struct btrfs_block_rsv *block_rsv;
6369         struct extent_buffer *buf;
6370         u64 flags = 0;
6371         int ret;
6372
6373
6374         block_rsv = use_block_rsv(trans, root, blocksize);
6375         if (IS_ERR(block_rsv))
6376                 return ERR_CAST(block_rsv);
6377
6378         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6379                                    empty_size, hint, &ins, 0);
6380         if (ret) {
6381                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6382                 return ERR_PTR(ret);
6383         }
6384
6385         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6386                                     blocksize, level);
6387         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6388
6389         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6390                 if (parent == 0)
6391                         parent = ins.objectid;
6392                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6393         } else
6394                 BUG_ON(parent > 0);
6395
6396         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6397                 struct btrfs_delayed_extent_op *extent_op;
6398                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6399                 BUG_ON(!extent_op); /* -ENOMEM */
6400                 if (key)
6401                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6402                 else
6403                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6404                 extent_op->flags_to_set = flags;
6405                 extent_op->update_key = 1;
6406                 extent_op->update_flags = 1;
6407                 extent_op->is_data = 0;
6408
6409                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6410                                         ins.objectid,
6411                                         ins.offset, parent, root_objectid,
6412                                         level, BTRFS_ADD_DELAYED_EXTENT,
6413                                         extent_op, 0);
6414                 BUG_ON(ret); /* -ENOMEM */
6415         }
6416         return buf;
6417 }
6418
6419 struct walk_control {
6420         u64 refs[BTRFS_MAX_LEVEL];
6421         u64 flags[BTRFS_MAX_LEVEL];
6422         struct btrfs_key update_progress;
6423         int stage;
6424         int level;
6425         int shared_level;
6426         int update_ref;
6427         int keep_locks;
6428         int reada_slot;
6429         int reada_count;
6430         int for_reloc;
6431 };
6432
6433 #define DROP_REFERENCE  1
6434 #define UPDATE_BACKREF  2
6435
6436 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6437                                      struct btrfs_root *root,
6438                                      struct walk_control *wc,
6439                                      struct btrfs_path *path)
6440 {
6441         u64 bytenr;
6442         u64 generation;
6443         u64 refs;
6444         u64 flags;
6445         u32 nritems;
6446         u32 blocksize;
6447         struct btrfs_key key;
6448         struct extent_buffer *eb;
6449         int ret;
6450         int slot;
6451         int nread = 0;
6452
6453         if (path->slots[wc->level] < wc->reada_slot) {
6454                 wc->reada_count = wc->reada_count * 2 / 3;
6455                 wc->reada_count = max(wc->reada_count, 2);
6456         } else {
6457                 wc->reada_count = wc->reada_count * 3 / 2;
6458                 wc->reada_count = min_t(int, wc->reada_count,
6459                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6460         }
6461
6462         eb = path->nodes[wc->level];
6463         nritems = btrfs_header_nritems(eb);
6464         blocksize = btrfs_level_size(root, wc->level - 1);
6465
6466         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6467                 if (nread >= wc->reada_count)
6468                         break;
6469
6470                 cond_resched();
6471                 bytenr = btrfs_node_blockptr(eb, slot);
6472                 generation = btrfs_node_ptr_generation(eb, slot);
6473
6474                 if (slot == path->slots[wc->level])
6475                         goto reada;
6476
6477                 if (wc->stage == UPDATE_BACKREF &&
6478                     generation <= root->root_key.offset)
6479                         continue;
6480
6481                 /* We don't lock the tree block, it's OK to be racy here */
6482                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6483                                                &refs, &flags);
6484                 /* We don't care about errors in readahead. */
6485                 if (ret < 0)
6486                         continue;
6487                 BUG_ON(refs == 0);
6488
6489                 if (wc->stage == DROP_REFERENCE) {
6490                         if (refs == 1)
6491                                 goto reada;
6492
6493                         if (wc->level == 1 &&
6494                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6495                                 continue;
6496                         if (!wc->update_ref ||
6497                             generation <= root->root_key.offset)
6498                                 continue;
6499                         btrfs_node_key_to_cpu(eb, &key, slot);
6500                         ret = btrfs_comp_cpu_keys(&key,
6501                                                   &wc->update_progress);
6502                         if (ret < 0)
6503                                 continue;
6504                 } else {
6505                         if (wc->level == 1 &&
6506                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6507                                 continue;
6508                 }
6509 reada:
6510                 ret = readahead_tree_block(root, bytenr, blocksize,
6511                                            generation);
6512                 if (ret)
6513                         break;
6514                 nread++;
6515         }
6516         wc->reada_slot = slot;
6517 }
6518
6519 /*
6520  * hepler to process tree block while walking down the tree.
6521  *
6522  * when wc->stage == UPDATE_BACKREF, this function updates
6523  * back refs for pointers in the block.
6524  *
6525  * NOTE: return value 1 means we should stop walking down.
6526  */
6527 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6528                                    struct btrfs_root *root,
6529                                    struct btrfs_path *path,
6530                                    struct walk_control *wc, int lookup_info)
6531 {
6532         int level = wc->level;
6533         struct extent_buffer *eb = path->nodes[level];
6534         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6535         int ret;
6536
6537         if (wc->stage == UPDATE_BACKREF &&
6538             btrfs_header_owner(eb) != root->root_key.objectid)
6539                 return 1;
6540
6541         /*
6542          * when reference count of tree block is 1, it won't increase
6543          * again. once full backref flag is set, we never clear it.
6544          */
6545         if (lookup_info &&
6546             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6547              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6548                 BUG_ON(!path->locks[level]);
6549                 ret = btrfs_lookup_extent_info(trans, root,
6550                                                eb->start, eb->len,
6551                                                &wc->refs[level],
6552                                                &wc->flags[level]);
6553                 BUG_ON(ret == -ENOMEM);
6554                 if (ret)
6555                         return ret;
6556                 BUG_ON(wc->refs[level] == 0);
6557         }
6558
6559         if (wc->stage == DROP_REFERENCE) {
6560                 if (wc->refs[level] > 1)
6561                         return 1;
6562
6563                 if (path->locks[level] && !wc->keep_locks) {
6564                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6565                         path->locks[level] = 0;
6566                 }
6567                 return 0;
6568         }
6569
6570         /* wc->stage == UPDATE_BACKREF */
6571         if (!(wc->flags[level] & flag)) {
6572                 BUG_ON(!path->locks[level]);
6573                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6574                 BUG_ON(ret); /* -ENOMEM */
6575                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6576                 BUG_ON(ret); /* -ENOMEM */
6577                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6578                                                   eb->len, flag, 0);
6579                 BUG_ON(ret); /* -ENOMEM */
6580                 wc->flags[level] |= flag;
6581         }
6582
6583         /*
6584          * the block is shared by multiple trees, so it's not good to
6585          * keep the tree lock
6586          */
6587         if (path->locks[level] && level > 0) {
6588                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6589                 path->locks[level] = 0;
6590         }
6591         return 0;
6592 }
6593
6594 /*
6595  * hepler to process tree block pointer.
6596  *
6597  * when wc->stage == DROP_REFERENCE, this function checks
6598  * reference count of the block pointed to. if the block
6599  * is shared and we need update back refs for the subtree
6600  * rooted at the block, this function changes wc->stage to
6601  * UPDATE_BACKREF. if the block is shared and there is no
6602  * need to update back, this function drops the reference
6603  * to the block.
6604  *
6605  * NOTE: return value 1 means we should stop walking down.
6606  */
6607 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6608                                  struct btrfs_root *root,
6609                                  struct btrfs_path *path,
6610                                  struct walk_control *wc, int *lookup_info)
6611 {
6612         u64 bytenr;
6613         u64 generation;
6614         u64 parent;
6615         u32 blocksize;
6616         struct btrfs_key key;
6617         struct extent_buffer *next;
6618         int level = wc->level;
6619         int reada = 0;
6620         int ret = 0;
6621
6622         generation = btrfs_node_ptr_generation(path->nodes[level],
6623                                                path->slots[level]);
6624         /*
6625          * if the lower level block was created before the snapshot
6626          * was created, we know there is no need to update back refs
6627          * for the subtree
6628          */
6629         if (wc->stage == UPDATE_BACKREF &&
6630             generation <= root->root_key.offset) {
6631                 *lookup_info = 1;
6632                 return 1;
6633         }
6634
6635         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6636         blocksize = btrfs_level_size(root, level - 1);
6637
6638         next = btrfs_find_tree_block(root, bytenr, blocksize);
6639         if (!next) {
6640                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6641                 if (!next)
6642                         return -ENOMEM;
6643                 reada = 1;
6644         }
6645         btrfs_tree_lock(next);
6646         btrfs_set_lock_blocking(next);
6647
6648         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6649                                        &wc->refs[level - 1],
6650                                        &wc->flags[level - 1]);
6651         if (ret < 0) {
6652                 btrfs_tree_unlock(next);
6653                 return ret;
6654         }
6655
6656         BUG_ON(wc->refs[level - 1] == 0);
6657         *lookup_info = 0;
6658
6659         if (wc->stage == DROP_REFERENCE) {
6660                 if (wc->refs[level - 1] > 1) {
6661                         if (level == 1 &&
6662                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6663                                 goto skip;
6664
6665                         if (!wc->update_ref ||
6666                             generation <= root->root_key.offset)
6667                                 goto skip;
6668
6669                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6670                                               path->slots[level]);
6671                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6672                         if (ret < 0)
6673                                 goto skip;
6674
6675                         wc->stage = UPDATE_BACKREF;
6676                         wc->shared_level = level - 1;
6677                 }
6678         } else {
6679                 if (level == 1 &&
6680                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6681                         goto skip;
6682         }
6683
6684         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6685                 btrfs_tree_unlock(next);
6686                 free_extent_buffer(next);
6687                 next = NULL;
6688                 *lookup_info = 1;
6689         }
6690
6691         if (!next) {
6692                 if (reada && level == 1)
6693                         reada_walk_down(trans, root, wc, path);
6694                 next = read_tree_block(root, bytenr, blocksize, generation);
6695                 if (!next)
6696                         return -EIO;
6697                 btrfs_tree_lock(next);
6698                 btrfs_set_lock_blocking(next);
6699         }
6700
6701         level--;
6702         BUG_ON(level != btrfs_header_level(next));
6703         path->nodes[level] = next;
6704         path->slots[level] = 0;
6705         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6706         wc->level = level;
6707         if (wc->level == 1)
6708                 wc->reada_slot = 0;
6709         return 0;
6710 skip:
6711         wc->refs[level - 1] = 0;
6712         wc->flags[level - 1] = 0;
6713         if (wc->stage == DROP_REFERENCE) {
6714                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6715                         parent = path->nodes[level]->start;
6716                 } else {
6717                         BUG_ON(root->root_key.objectid !=
6718                                btrfs_header_owner(path->nodes[level]));
6719                         parent = 0;
6720                 }
6721
6722                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6723                                 root->root_key.objectid, level - 1, 0, 0);
6724                 BUG_ON(ret); /* -ENOMEM */
6725         }
6726         btrfs_tree_unlock(next);
6727         free_extent_buffer(next);
6728         *lookup_info = 1;
6729         return 1;
6730 }
6731
6732 /*
6733  * hepler to process tree block while walking up the tree.
6734  *
6735  * when wc->stage == DROP_REFERENCE, this function drops
6736  * reference count on the block.
6737  *
6738  * when wc->stage == UPDATE_BACKREF, this function changes
6739  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6740  * to UPDATE_BACKREF previously while processing the block.
6741  *
6742  * NOTE: return value 1 means we should stop walking up.
6743  */
6744 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6745                                  struct btrfs_root *root,
6746                                  struct btrfs_path *path,
6747                                  struct walk_control *wc)
6748 {
6749         int ret;
6750         int level = wc->level;
6751         struct extent_buffer *eb = path->nodes[level];
6752         u64 parent = 0;
6753
6754         if (wc->stage == UPDATE_BACKREF) {
6755                 BUG_ON(wc->shared_level < level);
6756                 if (level < wc->shared_level)
6757                         goto out;
6758
6759                 ret = find_next_key(path, level + 1, &wc->update_progress);
6760                 if (ret > 0)
6761                         wc->update_ref = 0;
6762
6763                 wc->stage = DROP_REFERENCE;
6764                 wc->shared_level = -1;
6765                 path->slots[level] = 0;
6766
6767                 /*
6768                  * check reference count again if the block isn't locked.
6769                  * we should start walking down the tree again if reference
6770                  * count is one.
6771                  */
6772                 if (!path->locks[level]) {
6773                         BUG_ON(level == 0);
6774                         btrfs_tree_lock(eb);
6775                         btrfs_set_lock_blocking(eb);
6776                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6777
6778                         ret = btrfs_lookup_extent_info(trans, root,
6779                                                        eb->start, eb->len,
6780                                                        &wc->refs[level],
6781                                                        &wc->flags[level]);
6782                         if (ret < 0) {
6783                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6784                                 return ret;
6785                         }
6786                         BUG_ON(wc->refs[level] == 0);
6787                         if (wc->refs[level] == 1) {
6788                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6789                                 return 1;
6790                         }
6791                 }
6792         }
6793
6794         /* wc->stage == DROP_REFERENCE */
6795         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6796
6797         if (wc->refs[level] == 1) {
6798                 if (level == 0) {
6799                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6800                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6801                                                     wc->for_reloc);
6802                         else
6803                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6804                                                     wc->for_reloc);
6805                         BUG_ON(ret); /* -ENOMEM */
6806                 }
6807                 /* make block locked assertion in clean_tree_block happy */
6808                 if (!path->locks[level] &&
6809                     btrfs_header_generation(eb) == trans->transid) {
6810                         btrfs_tree_lock(eb);
6811                         btrfs_set_lock_blocking(eb);
6812                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6813                 }
6814                 clean_tree_block(trans, root, eb);
6815         }
6816
6817         if (eb == root->node) {
6818                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6819                         parent = eb->start;
6820                 else
6821                         BUG_ON(root->root_key.objectid !=
6822                                btrfs_header_owner(eb));
6823         } else {
6824                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6825                         parent = path->nodes[level + 1]->start;
6826                 else
6827                         BUG_ON(root->root_key.objectid !=
6828                                btrfs_header_owner(path->nodes[level + 1]));
6829         }
6830
6831         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6832 out:
6833         wc->refs[level] = 0;
6834         wc->flags[level] = 0;
6835         return 0;
6836 }
6837
6838 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6839                                    struct btrfs_root *root,
6840                                    struct btrfs_path *path,
6841                                    struct walk_control *wc)
6842 {
6843         int level = wc->level;
6844         int lookup_info = 1;
6845         int ret;
6846
6847         while (level >= 0) {
6848                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6849                 if (ret > 0)
6850                         break;
6851
6852                 if (level == 0)
6853                         break;
6854
6855                 if (path->slots[level] >=
6856                     btrfs_header_nritems(path->nodes[level]))
6857                         break;
6858
6859                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6860                 if (ret > 0) {
6861                         path->slots[level]++;
6862                         continue;
6863                 } else if (ret < 0)
6864                         return ret;
6865                 level = wc->level;
6866         }
6867         return 0;
6868 }
6869
6870 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6871                                  struct btrfs_root *root,
6872                                  struct btrfs_path *path,
6873                                  struct walk_control *wc, int max_level)
6874 {
6875         int level = wc->level;
6876         int ret;
6877
6878         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6879         while (level < max_level && path->nodes[level]) {
6880                 wc->level = level;
6881                 if (path->slots[level] + 1 <
6882                     btrfs_header_nritems(path->nodes[level])) {
6883                         path->slots[level]++;
6884                         return 0;
6885                 } else {
6886                         ret = walk_up_proc(trans, root, path, wc);
6887                         if (ret > 0)
6888                                 return 0;
6889
6890                         if (path->locks[level]) {
6891                                 btrfs_tree_unlock_rw(path->nodes[level],
6892                                                      path->locks[level]);
6893                                 path->locks[level] = 0;
6894                         }
6895                         free_extent_buffer(path->nodes[level]);
6896                         path->nodes[level] = NULL;
6897                         level++;
6898                 }
6899         }
6900         return 1;
6901 }
6902
6903 /*
6904  * drop a subvolume tree.
6905  *
6906  * this function traverses the tree freeing any blocks that only
6907  * referenced by the tree.
6908  *
6909  * when a shared tree block is found. this function decreases its
6910  * reference count by one. if update_ref is true, this function
6911  * also make sure backrefs for the shared block and all lower level
6912  * blocks are properly updated.
6913  */
6914 int btrfs_drop_snapshot(struct btrfs_root *root,
6915                          struct btrfs_block_rsv *block_rsv, int update_ref,
6916                          int for_reloc)
6917 {
6918         struct btrfs_path *path;
6919         struct btrfs_trans_handle *trans;
6920         struct btrfs_root *tree_root = root->fs_info->tree_root;
6921         struct btrfs_root_item *root_item = &root->root_item;
6922         struct walk_control *wc;
6923         struct btrfs_key key;
6924         int err = 0;
6925         int ret;
6926         int level;
6927
6928         path = btrfs_alloc_path();
6929         if (!path) {
6930                 err = -ENOMEM;
6931                 goto out;
6932         }
6933
6934         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6935         if (!wc) {
6936                 btrfs_free_path(path);
6937                 err = -ENOMEM;
6938                 goto out;
6939         }
6940
6941         trans = btrfs_start_transaction(tree_root, 0);
6942         if (IS_ERR(trans)) {
6943                 err = PTR_ERR(trans);
6944                 goto out_free;
6945         }
6946
6947         if (block_rsv)
6948                 trans->block_rsv = block_rsv;
6949
6950         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6951                 level = btrfs_header_level(root->node);
6952                 path->nodes[level] = btrfs_lock_root_node(root);
6953                 btrfs_set_lock_blocking(path->nodes[level]);
6954                 path->slots[level] = 0;
6955                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6956                 memset(&wc->update_progress, 0,
6957                        sizeof(wc->update_progress));
6958         } else {
6959                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6960                 memcpy(&wc->update_progress, &key,
6961                        sizeof(wc->update_progress));
6962
6963                 level = root_item->drop_level;
6964                 BUG_ON(level == 0);
6965                 path->lowest_level = level;
6966                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6967                 path->lowest_level = 0;
6968                 if (ret < 0) {
6969                         err = ret;
6970                         goto out_end_trans;
6971                 }
6972                 WARN_ON(ret > 0);
6973
6974                 /*
6975                  * unlock our path, this is safe because only this
6976                  * function is allowed to delete this snapshot
6977                  */
6978                 btrfs_unlock_up_safe(path, 0);
6979
6980                 level = btrfs_header_level(root->node);
6981                 while (1) {
6982                         btrfs_tree_lock(path->nodes[level]);
6983                         btrfs_set_lock_blocking(path->nodes[level]);
6984
6985                         ret = btrfs_lookup_extent_info(trans, root,
6986                                                 path->nodes[level]->start,
6987                                                 path->nodes[level]->len,
6988                                                 &wc->refs[level],
6989                                                 &wc->flags[level]);
6990                         if (ret < 0) {
6991                                 err = ret;
6992                                 goto out_end_trans;
6993                         }
6994                         BUG_ON(wc->refs[level] == 0);
6995
6996                         if (level == root_item->drop_level)
6997                                 break;
6998
6999                         btrfs_tree_unlock(path->nodes[level]);
7000                         WARN_ON(wc->refs[level] != 1);
7001                         level--;
7002                 }
7003         }
7004
7005         wc->level = level;
7006         wc->shared_level = -1;
7007         wc->stage = DROP_REFERENCE;
7008         wc->update_ref = update_ref;
7009         wc->keep_locks = 0;
7010         wc->for_reloc = for_reloc;
7011         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7012
7013         while (1) {
7014                 ret = walk_down_tree(trans, root, path, wc);
7015                 if (ret < 0) {
7016                         err = ret;
7017                         break;
7018                 }
7019
7020                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7021                 if (ret < 0) {
7022                         err = ret;
7023                         break;
7024                 }
7025
7026                 if (ret > 0) {
7027                         BUG_ON(wc->stage != DROP_REFERENCE);
7028                         break;
7029                 }
7030
7031                 if (wc->stage == DROP_REFERENCE) {
7032                         level = wc->level;
7033                         btrfs_node_key(path->nodes[level],
7034                                        &root_item->drop_progress,
7035                                        path->slots[level]);
7036                         root_item->drop_level = level;
7037                 }
7038
7039                 BUG_ON(wc->level == 0);
7040                 if (btrfs_should_end_transaction(trans, tree_root)) {
7041                         ret = btrfs_update_root(trans, tree_root,
7042                                                 &root->root_key,
7043                                                 root_item);
7044                         if (ret) {
7045                                 btrfs_abort_transaction(trans, tree_root, ret);
7046                                 err = ret;
7047                                 goto out_end_trans;
7048                         }
7049
7050                         btrfs_end_transaction_throttle(trans, tree_root);
7051                         trans = btrfs_start_transaction(tree_root, 0);
7052                         if (IS_ERR(trans)) {
7053                                 err = PTR_ERR(trans);
7054                                 goto out_free;
7055                         }
7056                         if (block_rsv)
7057                                 trans->block_rsv = block_rsv;
7058                 }
7059         }
7060         btrfs_release_path(path);
7061         if (err)
7062                 goto out_end_trans;
7063
7064         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7065         if (ret) {
7066                 btrfs_abort_transaction(trans, tree_root, ret);
7067                 goto out_end_trans;
7068         }
7069
7070         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7071                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7072                                            NULL, NULL);
7073                 if (ret < 0) {
7074                         btrfs_abort_transaction(trans, tree_root, ret);
7075                         err = ret;
7076                         goto out_end_trans;
7077                 } else if (ret > 0) {
7078                         /* if we fail to delete the orphan item this time
7079                          * around, it'll get picked up the next time.
7080                          *
7081                          * The most common failure here is just -ENOENT.
7082                          */
7083                         btrfs_del_orphan_item(trans, tree_root,
7084                                               root->root_key.objectid);
7085                 }
7086         }
7087
7088         if (root->in_radix) {
7089                 btrfs_free_fs_root(tree_root->fs_info, root);
7090         } else {
7091                 free_extent_buffer(root->node);
7092                 free_extent_buffer(root->commit_root);
7093                 kfree(root);
7094         }
7095 out_end_trans:
7096         btrfs_end_transaction_throttle(trans, tree_root);
7097 out_free:
7098         kfree(wc);
7099         btrfs_free_path(path);
7100 out:
7101         if (err)
7102                 btrfs_std_error(root->fs_info, err);
7103         return err;
7104 }
7105
7106 /*
7107  * drop subtree rooted at tree block 'node'.
7108  *
7109  * NOTE: this function will unlock and release tree block 'node'
7110  * only used by relocation code
7111  */
7112 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7113                         struct btrfs_root *root,
7114                         struct extent_buffer *node,
7115                         struct extent_buffer *parent)
7116 {
7117         struct btrfs_path *path;
7118         struct walk_control *wc;
7119         int level;
7120         int parent_level;
7121         int ret = 0;
7122         int wret;
7123
7124         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7125
7126         path = btrfs_alloc_path();
7127         if (!path)
7128                 return -ENOMEM;
7129
7130         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7131         if (!wc) {
7132                 btrfs_free_path(path);
7133                 return -ENOMEM;
7134         }
7135
7136         btrfs_assert_tree_locked(parent);
7137         parent_level = btrfs_header_level(parent);
7138         extent_buffer_get(parent);
7139         path->nodes[parent_level] = parent;
7140         path->slots[parent_level] = btrfs_header_nritems(parent);
7141
7142         btrfs_assert_tree_locked(node);
7143         level = btrfs_header_level(node);
7144         path->nodes[level] = node;
7145         path->slots[level] = 0;
7146         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7147
7148         wc->refs[parent_level] = 1;
7149         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7150         wc->level = level;
7151         wc->shared_level = -1;
7152         wc->stage = DROP_REFERENCE;
7153         wc->update_ref = 0;
7154         wc->keep_locks = 1;
7155         wc->for_reloc = 1;
7156         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7157
7158         while (1) {
7159                 wret = walk_down_tree(trans, root, path, wc);
7160                 if (wret < 0) {
7161                         ret = wret;
7162                         break;
7163                 }
7164
7165                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7166                 if (wret < 0)
7167                         ret = wret;
7168                 if (wret != 0)
7169                         break;
7170         }
7171
7172         kfree(wc);
7173         btrfs_free_path(path);
7174         return ret;
7175 }
7176
7177 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7178 {
7179         u64 num_devices;
7180         u64 stripped;
7181
7182         /*
7183          * if restripe for this chunk_type is on pick target profile and
7184          * return, otherwise do the usual balance
7185          */
7186         stripped = get_restripe_target(root->fs_info, flags);
7187         if (stripped)
7188                 return extended_to_chunk(stripped);
7189
7190         /*
7191          * we add in the count of missing devices because we want
7192          * to make sure that any RAID levels on a degraded FS
7193          * continue to be honored.
7194          */
7195         num_devices = root->fs_info->fs_devices->rw_devices +
7196                 root->fs_info->fs_devices->missing_devices;
7197
7198         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7199                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7200
7201         if (num_devices == 1) {
7202                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7203                 stripped = flags & ~stripped;
7204
7205                 /* turn raid0 into single device chunks */
7206                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7207                         return stripped;
7208
7209                 /* turn mirroring into duplication */
7210                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7211                              BTRFS_BLOCK_GROUP_RAID10))
7212                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7213         } else {
7214                 /* they already had raid on here, just return */
7215                 if (flags & stripped)
7216                         return flags;
7217
7218                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7219                 stripped = flags & ~stripped;
7220
7221                 /* switch duplicated blocks with raid1 */
7222                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7223                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7224
7225                 /* this is drive concat, leave it alone */
7226         }
7227
7228         return flags;
7229 }
7230
7231 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7232 {
7233         struct btrfs_space_info *sinfo = cache->space_info;
7234         u64 num_bytes;
7235         u64 min_allocable_bytes;
7236         int ret = -ENOSPC;
7237
7238
7239         /*
7240          * We need some metadata space and system metadata space for
7241          * allocating chunks in some corner cases until we force to set
7242          * it to be readonly.
7243          */
7244         if ((sinfo->flags &
7245              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7246             !force)
7247                 min_allocable_bytes = 1 * 1024 * 1024;
7248         else
7249                 min_allocable_bytes = 0;
7250
7251         spin_lock(&sinfo->lock);
7252         spin_lock(&cache->lock);
7253
7254         if (cache->ro) {
7255                 ret = 0;
7256                 goto out;
7257         }
7258
7259         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7260                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7261
7262         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7263             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7264             min_allocable_bytes <= sinfo->total_bytes) {
7265                 sinfo->bytes_readonly += num_bytes;
7266                 cache->ro = 1;
7267                 ret = 0;
7268         }
7269 out:
7270         spin_unlock(&cache->lock);
7271         spin_unlock(&sinfo->lock);
7272         return ret;
7273 }
7274
7275 int btrfs_set_block_group_ro(struct btrfs_root *root,
7276                              struct btrfs_block_group_cache *cache)
7277
7278 {
7279         struct btrfs_trans_handle *trans;
7280         u64 alloc_flags;
7281         int ret;
7282
7283         BUG_ON(cache->ro);
7284
7285         trans = btrfs_join_transaction(root);
7286         if (IS_ERR(trans))
7287                 return PTR_ERR(trans);
7288
7289         alloc_flags = update_block_group_flags(root, cache->flags);
7290         if (alloc_flags != cache->flags) {
7291                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7292                                      CHUNK_ALLOC_FORCE);
7293                 if (ret < 0)
7294                         goto out;
7295         }
7296
7297         ret = set_block_group_ro(cache, 0);
7298         if (!ret)
7299                 goto out;
7300         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7301         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7302                              CHUNK_ALLOC_FORCE);
7303         if (ret < 0)
7304                 goto out;
7305         ret = set_block_group_ro(cache, 0);
7306 out:
7307         btrfs_end_transaction(trans, root);
7308         return ret;
7309 }
7310
7311 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7312                             struct btrfs_root *root, u64 type)
7313 {
7314         u64 alloc_flags = get_alloc_profile(root, type);
7315         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7316                               CHUNK_ALLOC_FORCE);
7317 }
7318
7319 /*
7320  * helper to account the unused space of all the readonly block group in the
7321  * list. takes mirrors into account.
7322  */
7323 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7324 {
7325         struct btrfs_block_group_cache *block_group;
7326         u64 free_bytes = 0;
7327         int factor;
7328
7329         list_for_each_entry(block_group, groups_list, list) {
7330                 spin_lock(&block_group->lock);
7331
7332                 if (!block_group->ro) {
7333                         spin_unlock(&block_group->lock);
7334                         continue;
7335                 }
7336
7337                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7338                                           BTRFS_BLOCK_GROUP_RAID10 |
7339                                           BTRFS_BLOCK_GROUP_DUP))
7340                         factor = 2;
7341                 else
7342                         factor = 1;
7343
7344                 free_bytes += (block_group->key.offset -
7345                                btrfs_block_group_used(&block_group->item)) *
7346                                factor;
7347
7348                 spin_unlock(&block_group->lock);
7349         }
7350
7351         return free_bytes;
7352 }
7353
7354 /*
7355  * helper to account the unused space of all the readonly block group in the
7356  * space_info. takes mirrors into account.
7357  */
7358 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7359 {
7360         int i;
7361         u64 free_bytes = 0;
7362
7363         spin_lock(&sinfo->lock);
7364
7365         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7366                 if (!list_empty(&sinfo->block_groups[i]))
7367                         free_bytes += __btrfs_get_ro_block_group_free_space(
7368                                                 &sinfo->block_groups[i]);
7369
7370         spin_unlock(&sinfo->lock);
7371
7372         return free_bytes;
7373 }
7374
7375 void btrfs_set_block_group_rw(struct btrfs_root *root,
7376                               struct btrfs_block_group_cache *cache)
7377 {
7378         struct btrfs_space_info *sinfo = cache->space_info;
7379         u64 num_bytes;
7380
7381         BUG_ON(!cache->ro);
7382
7383         spin_lock(&sinfo->lock);
7384         spin_lock(&cache->lock);
7385         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7386                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7387         sinfo->bytes_readonly -= num_bytes;
7388         cache->ro = 0;
7389         spin_unlock(&cache->lock);
7390         spin_unlock(&sinfo->lock);
7391 }
7392
7393 /*
7394  * checks to see if its even possible to relocate this block group.
7395  *
7396  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7397  * ok to go ahead and try.
7398  */
7399 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7400 {
7401         struct btrfs_block_group_cache *block_group;
7402         struct btrfs_space_info *space_info;
7403         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7404         struct btrfs_device *device;
7405         u64 min_free;
7406         u64 dev_min = 1;
7407         u64 dev_nr = 0;
7408         u64 target;
7409         int index;
7410         int full = 0;
7411         int ret = 0;
7412
7413         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7414
7415         /* odd, couldn't find the block group, leave it alone */
7416         if (!block_group)
7417                 return -1;
7418
7419         min_free = btrfs_block_group_used(&block_group->item);
7420
7421         /* no bytes used, we're good */
7422         if (!min_free)
7423                 goto out;
7424
7425         space_info = block_group->space_info;
7426         spin_lock(&space_info->lock);
7427
7428         full = space_info->full;
7429
7430         /*
7431          * if this is the last block group we have in this space, we can't
7432          * relocate it unless we're able to allocate a new chunk below.
7433          *
7434          * Otherwise, we need to make sure we have room in the space to handle
7435          * all of the extents from this block group.  If we can, we're good
7436          */
7437         if ((space_info->total_bytes != block_group->key.offset) &&
7438             (space_info->bytes_used + space_info->bytes_reserved +
7439              space_info->bytes_pinned + space_info->bytes_readonly +
7440              min_free < space_info->total_bytes)) {
7441                 spin_unlock(&space_info->lock);
7442                 goto out;
7443         }
7444         spin_unlock(&space_info->lock);
7445
7446         /*
7447          * ok we don't have enough space, but maybe we have free space on our
7448          * devices to allocate new chunks for relocation, so loop through our
7449          * alloc devices and guess if we have enough space.  if this block
7450          * group is going to be restriped, run checks against the target
7451          * profile instead of the current one.
7452          */
7453         ret = -1;
7454
7455         /*
7456          * index:
7457          *      0: raid10
7458          *      1: raid1
7459          *      2: dup
7460          *      3: raid0
7461          *      4: single
7462          */
7463         target = get_restripe_target(root->fs_info, block_group->flags);
7464         if (target) {
7465                 index = __get_block_group_index(extended_to_chunk(target));
7466         } else {
7467                 /*
7468                  * this is just a balance, so if we were marked as full
7469                  * we know there is no space for a new chunk
7470                  */
7471                 if (full)
7472                         goto out;
7473
7474                 index = get_block_group_index(block_group);
7475         }
7476
7477         if (index == 0) {
7478                 dev_min = 4;
7479                 /* Divide by 2 */
7480                 min_free >>= 1;
7481         } else if (index == 1) {
7482                 dev_min = 2;
7483         } else if (index == 2) {
7484                 /* Multiply by 2 */
7485                 min_free <<= 1;
7486         } else if (index == 3) {
7487                 dev_min = fs_devices->rw_devices;
7488                 do_div(min_free, dev_min);
7489         }
7490
7491         mutex_lock(&root->fs_info->chunk_mutex);
7492         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7493                 u64 dev_offset;
7494
7495                 /*
7496                  * check to make sure we can actually find a chunk with enough
7497                  * space to fit our block group in.
7498                  */
7499                 if (device->total_bytes > device->bytes_used + min_free) {
7500                         ret = find_free_dev_extent(device, min_free,
7501                                                    &dev_offset, NULL);
7502                         if (!ret)
7503                                 dev_nr++;
7504
7505                         if (dev_nr >= dev_min)
7506                                 break;
7507
7508                         ret = -1;
7509                 }
7510         }
7511         mutex_unlock(&root->fs_info->chunk_mutex);
7512 out:
7513         btrfs_put_block_group(block_group);
7514         return ret;
7515 }
7516
7517 static int find_first_block_group(struct btrfs_root *root,
7518                 struct btrfs_path *path, struct btrfs_key *key)
7519 {
7520         int ret = 0;
7521         struct btrfs_key found_key;
7522         struct extent_buffer *leaf;
7523         int slot;
7524
7525         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7526         if (ret < 0)
7527                 goto out;
7528
7529         while (1) {
7530                 slot = path->slots[0];
7531                 leaf = path->nodes[0];
7532                 if (slot >= btrfs_header_nritems(leaf)) {
7533                         ret = btrfs_next_leaf(root, path);
7534                         if (ret == 0)
7535                                 continue;
7536                         if (ret < 0)
7537                                 goto out;
7538                         break;
7539                 }
7540                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7541
7542                 if (found_key.objectid >= key->objectid &&
7543                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7544                         ret = 0;
7545                         goto out;
7546                 }
7547                 path->slots[0]++;
7548         }
7549 out:
7550         return ret;
7551 }
7552
7553 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7554 {
7555         struct btrfs_block_group_cache *block_group;
7556         u64 last = 0;
7557
7558         while (1) {
7559                 struct inode *inode;
7560
7561                 block_group = btrfs_lookup_first_block_group(info, last);
7562                 while (block_group) {
7563                         spin_lock(&block_group->lock);
7564                         if (block_group->iref)
7565                                 break;
7566                         spin_unlock(&block_group->lock);
7567                         block_group = next_block_group(info->tree_root,
7568                                                        block_group);
7569                 }
7570                 if (!block_group) {
7571                         if (last == 0)
7572                                 break;
7573                         last = 0;
7574                         continue;
7575                 }
7576
7577                 inode = block_group->inode;
7578                 block_group->iref = 0;
7579                 block_group->inode = NULL;
7580                 spin_unlock(&block_group->lock);
7581                 iput(inode);
7582                 last = block_group->key.objectid + block_group->key.offset;
7583                 btrfs_put_block_group(block_group);
7584         }
7585 }
7586
7587 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7588 {
7589         struct btrfs_block_group_cache *block_group;
7590         struct btrfs_space_info *space_info;
7591         struct btrfs_caching_control *caching_ctl;
7592         struct rb_node *n;
7593
7594         down_write(&info->extent_commit_sem);
7595         while (!list_empty(&info->caching_block_groups)) {
7596                 caching_ctl = list_entry(info->caching_block_groups.next,
7597                                          struct btrfs_caching_control, list);
7598                 list_del(&caching_ctl->list);
7599                 put_caching_control(caching_ctl);
7600         }
7601         up_write(&info->extent_commit_sem);
7602
7603         spin_lock(&info->block_group_cache_lock);
7604         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7605                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7606                                        cache_node);
7607                 rb_erase(&block_group->cache_node,
7608                          &info->block_group_cache_tree);
7609                 spin_unlock(&info->block_group_cache_lock);
7610
7611                 down_write(&block_group->space_info->groups_sem);
7612                 list_del(&block_group->list);
7613                 up_write(&block_group->space_info->groups_sem);
7614
7615                 if (block_group->cached == BTRFS_CACHE_STARTED)
7616                         wait_block_group_cache_done(block_group);
7617
7618                 /*
7619                  * We haven't cached this block group, which means we could
7620                  * possibly have excluded extents on this block group.
7621                  */
7622                 if (block_group->cached == BTRFS_CACHE_NO)
7623                         free_excluded_extents(info->extent_root, block_group);
7624
7625                 btrfs_remove_free_space_cache(block_group);
7626                 btrfs_put_block_group(block_group);
7627
7628                 spin_lock(&info->block_group_cache_lock);
7629         }
7630         spin_unlock(&info->block_group_cache_lock);
7631
7632         /* now that all the block groups are freed, go through and
7633          * free all the space_info structs.  This is only called during
7634          * the final stages of unmount, and so we know nobody is
7635          * using them.  We call synchronize_rcu() once before we start,
7636          * just to be on the safe side.
7637          */
7638         synchronize_rcu();
7639
7640         release_global_block_rsv(info);
7641
7642         while(!list_empty(&info->space_info)) {
7643                 space_info = list_entry(info->space_info.next,
7644                                         struct btrfs_space_info,
7645                                         list);
7646                 if (space_info->bytes_pinned > 0 ||
7647                     space_info->bytes_reserved > 0 ||
7648                     space_info->bytes_may_use > 0) {
7649                         WARN_ON(1);
7650                         dump_space_info(space_info, 0, 0);
7651                 }
7652                 list_del(&space_info->list);
7653                 kfree(space_info);
7654         }
7655         return 0;
7656 }
7657
7658 static void __link_block_group(struct btrfs_space_info *space_info,
7659                                struct btrfs_block_group_cache *cache)
7660 {
7661         int index = get_block_group_index(cache);
7662
7663         down_write(&space_info->groups_sem);
7664         list_add_tail(&cache->list, &space_info->block_groups[index]);
7665         up_write(&space_info->groups_sem);
7666 }
7667
7668 int btrfs_read_block_groups(struct btrfs_root *root)
7669 {
7670         struct btrfs_path *path;
7671         int ret;
7672         struct btrfs_block_group_cache *cache;
7673         struct btrfs_fs_info *info = root->fs_info;
7674         struct btrfs_space_info *space_info;
7675         struct btrfs_key key;
7676         struct btrfs_key found_key;
7677         struct extent_buffer *leaf;
7678         int need_clear = 0;
7679         u64 cache_gen;
7680
7681         root = info->extent_root;
7682         key.objectid = 0;
7683         key.offset = 0;
7684         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7685         path = btrfs_alloc_path();
7686         if (!path)
7687                 return -ENOMEM;
7688         path->reada = 1;
7689
7690         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7691         if (btrfs_test_opt(root, SPACE_CACHE) &&
7692             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7693                 need_clear = 1;
7694         if (btrfs_test_opt(root, CLEAR_CACHE))
7695                 need_clear = 1;
7696
7697         while (1) {
7698                 ret = find_first_block_group(root, path, &key);
7699                 if (ret > 0)
7700                         break;
7701                 if (ret != 0)
7702                         goto error;
7703                 leaf = path->nodes[0];
7704                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7705                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7706                 if (!cache) {
7707                         ret = -ENOMEM;
7708                         goto error;
7709                 }
7710                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7711                                                 GFP_NOFS);
7712                 if (!cache->free_space_ctl) {
7713                         kfree(cache);
7714                         ret = -ENOMEM;
7715                         goto error;
7716                 }
7717
7718                 atomic_set(&cache->count, 1);
7719                 spin_lock_init(&cache->lock);
7720                 cache->fs_info = info;
7721                 INIT_LIST_HEAD(&cache->list);
7722                 INIT_LIST_HEAD(&cache->cluster_list);
7723
7724                 if (need_clear) {
7725                         /*
7726                          * When we mount with old space cache, we need to
7727                          * set BTRFS_DC_CLEAR and set dirty flag.
7728                          *
7729                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7730                          *    truncate the old free space cache inode and
7731                          *    setup a new one.
7732                          * b) Setting 'dirty flag' makes sure that we flush
7733                          *    the new space cache info onto disk.
7734                          */
7735                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7736                         if (btrfs_test_opt(root, SPACE_CACHE))
7737                                 cache->dirty = 1;
7738                 }
7739
7740                 read_extent_buffer(leaf, &cache->item,
7741                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7742                                    sizeof(cache->item));
7743                 memcpy(&cache->key, &found_key, sizeof(found_key));
7744
7745                 key.objectid = found_key.objectid + found_key.offset;
7746                 btrfs_release_path(path);
7747                 cache->flags = btrfs_block_group_flags(&cache->item);
7748                 cache->sectorsize = root->sectorsize;
7749
7750                 btrfs_init_free_space_ctl(cache);
7751
7752                 /*
7753                  * We need to exclude the super stripes now so that the space
7754                  * info has super bytes accounted for, otherwise we'll think
7755                  * we have more space than we actually do.
7756                  */
7757                 exclude_super_stripes(root, cache);
7758
7759                 /*
7760                  * check for two cases, either we are full, and therefore
7761                  * don't need to bother with the caching work since we won't
7762                  * find any space, or we are empty, and we can just add all
7763                  * the space in and be done with it.  This saves us _alot_ of
7764                  * time, particularly in the full case.
7765                  */
7766                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7767                         cache->last_byte_to_unpin = (u64)-1;
7768                         cache->cached = BTRFS_CACHE_FINISHED;
7769                         free_excluded_extents(root, cache);
7770                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7771                         cache->last_byte_to_unpin = (u64)-1;
7772                         cache->cached = BTRFS_CACHE_FINISHED;
7773                         add_new_free_space(cache, root->fs_info,
7774                                            found_key.objectid,
7775                                            found_key.objectid +
7776                                            found_key.offset);
7777                         free_excluded_extents(root, cache);
7778                 }
7779
7780                 ret = update_space_info(info, cache->flags, found_key.offset,
7781                                         btrfs_block_group_used(&cache->item),
7782                                         &space_info);
7783                 BUG_ON(ret); /* -ENOMEM */
7784                 cache->space_info = space_info;
7785                 spin_lock(&cache->space_info->lock);
7786                 cache->space_info->bytes_readonly += cache->bytes_super;
7787                 spin_unlock(&cache->space_info->lock);
7788
7789                 __link_block_group(space_info, cache);
7790
7791                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7792                 BUG_ON(ret); /* Logic error */
7793
7794                 set_avail_alloc_bits(root->fs_info, cache->flags);
7795                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7796                         set_block_group_ro(cache, 1);
7797         }
7798
7799         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7800                 if (!(get_alloc_profile(root, space_info->flags) &
7801                       (BTRFS_BLOCK_GROUP_RAID10 |
7802                        BTRFS_BLOCK_GROUP_RAID1 |
7803                        BTRFS_BLOCK_GROUP_DUP)))
7804                         continue;
7805                 /*
7806                  * avoid allocating from un-mirrored block group if there are
7807                  * mirrored block groups.
7808                  */
7809                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7810                         set_block_group_ro(cache, 1);
7811                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7812                         set_block_group_ro(cache, 1);
7813         }
7814
7815         init_global_block_rsv(info);
7816         ret = 0;
7817 error:
7818         btrfs_free_path(path);
7819         return ret;
7820 }
7821
7822 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7823                            struct btrfs_root *root, u64 bytes_used,
7824                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7825                            u64 size)
7826 {
7827         int ret;
7828         struct btrfs_root *extent_root;
7829         struct btrfs_block_group_cache *cache;
7830
7831         extent_root = root->fs_info->extent_root;
7832
7833         root->fs_info->last_trans_log_full_commit = trans->transid;
7834
7835         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7836         if (!cache)
7837                 return -ENOMEM;
7838         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7839                                         GFP_NOFS);
7840         if (!cache->free_space_ctl) {
7841                 kfree(cache);
7842                 return -ENOMEM;
7843         }
7844
7845         cache->key.objectid = chunk_offset;
7846         cache->key.offset = size;
7847         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7848         cache->sectorsize = root->sectorsize;
7849         cache->fs_info = root->fs_info;
7850
7851         atomic_set(&cache->count, 1);
7852         spin_lock_init(&cache->lock);
7853         INIT_LIST_HEAD(&cache->list);
7854         INIT_LIST_HEAD(&cache->cluster_list);
7855
7856         btrfs_init_free_space_ctl(cache);
7857
7858         btrfs_set_block_group_used(&cache->item, bytes_used);
7859         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7860         cache->flags = type;
7861         btrfs_set_block_group_flags(&cache->item, type);
7862
7863         cache->last_byte_to_unpin = (u64)-1;
7864         cache->cached = BTRFS_CACHE_FINISHED;
7865         exclude_super_stripes(root, cache);
7866
7867         add_new_free_space(cache, root->fs_info, chunk_offset,
7868                            chunk_offset + size);
7869
7870         free_excluded_extents(root, cache);
7871
7872         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7873                                 &cache->space_info);
7874         BUG_ON(ret); /* -ENOMEM */
7875         update_global_block_rsv(root->fs_info);
7876
7877         spin_lock(&cache->space_info->lock);
7878         cache->space_info->bytes_readonly += cache->bytes_super;
7879         spin_unlock(&cache->space_info->lock);
7880
7881         __link_block_group(cache->space_info, cache);
7882
7883         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7884         BUG_ON(ret); /* Logic error */
7885
7886         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7887                                 sizeof(cache->item));
7888         if (ret) {
7889                 btrfs_abort_transaction(trans, extent_root, ret);
7890                 return ret;
7891         }
7892
7893         set_avail_alloc_bits(extent_root->fs_info, type);
7894
7895         return 0;
7896 }
7897
7898 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7899 {
7900         u64 extra_flags = chunk_to_extended(flags) &
7901                                 BTRFS_EXTENDED_PROFILE_MASK;
7902
7903         if (flags & BTRFS_BLOCK_GROUP_DATA)
7904                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7905         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7906                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7907         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7908                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7909 }
7910
7911 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7912                              struct btrfs_root *root, u64 group_start)
7913 {
7914         struct btrfs_path *path;
7915         struct btrfs_block_group_cache *block_group;
7916         struct btrfs_free_cluster *cluster;
7917         struct btrfs_root *tree_root = root->fs_info->tree_root;
7918         struct btrfs_key key;
7919         struct inode *inode;
7920         int ret;
7921         int index;
7922         int factor;
7923
7924         root = root->fs_info->extent_root;
7925
7926         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7927         BUG_ON(!block_group);
7928         BUG_ON(!block_group->ro);
7929
7930         /*
7931          * Free the reserved super bytes from this block group before
7932          * remove it.
7933          */
7934         free_excluded_extents(root, block_group);
7935
7936         memcpy(&key, &block_group->key, sizeof(key));
7937         index = get_block_group_index(block_group);
7938         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7939                                   BTRFS_BLOCK_GROUP_RAID1 |
7940                                   BTRFS_BLOCK_GROUP_RAID10))
7941                 factor = 2;
7942         else
7943                 factor = 1;
7944
7945         /* make sure this block group isn't part of an allocation cluster */
7946         cluster = &root->fs_info->data_alloc_cluster;
7947         spin_lock(&cluster->refill_lock);
7948         btrfs_return_cluster_to_free_space(block_group, cluster);
7949         spin_unlock(&cluster->refill_lock);
7950
7951         /*
7952          * make sure this block group isn't part of a metadata
7953          * allocation cluster
7954          */
7955         cluster = &root->fs_info->meta_alloc_cluster;
7956         spin_lock(&cluster->refill_lock);
7957         btrfs_return_cluster_to_free_space(block_group, cluster);
7958         spin_unlock(&cluster->refill_lock);
7959
7960         path = btrfs_alloc_path();
7961         if (!path) {
7962                 ret = -ENOMEM;
7963                 goto out;
7964         }
7965
7966         inode = lookup_free_space_inode(tree_root, block_group, path);
7967         if (!IS_ERR(inode)) {
7968                 ret = btrfs_orphan_add(trans, inode);
7969                 if (ret) {
7970                         btrfs_add_delayed_iput(inode);
7971                         goto out;
7972                 }
7973                 clear_nlink(inode);
7974                 /* One for the block groups ref */
7975                 spin_lock(&block_group->lock);
7976                 if (block_group->iref) {
7977                         block_group->iref = 0;
7978                         block_group->inode = NULL;
7979                         spin_unlock(&block_group->lock);
7980                         iput(inode);
7981                 } else {
7982                         spin_unlock(&block_group->lock);
7983                 }
7984                 /* One for our lookup ref */
7985                 btrfs_add_delayed_iput(inode);
7986         }
7987
7988         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7989         key.offset = block_group->key.objectid;
7990         key.type = 0;
7991
7992         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7993         if (ret < 0)
7994                 goto out;
7995         if (ret > 0)
7996                 btrfs_release_path(path);
7997         if (ret == 0) {
7998                 ret = btrfs_del_item(trans, tree_root, path);
7999                 if (ret)
8000                         goto out;
8001                 btrfs_release_path(path);
8002         }
8003
8004         spin_lock(&root->fs_info->block_group_cache_lock);
8005         rb_erase(&block_group->cache_node,
8006                  &root->fs_info->block_group_cache_tree);
8007         spin_unlock(&root->fs_info->block_group_cache_lock);
8008
8009         down_write(&block_group->space_info->groups_sem);
8010         /*
8011          * we must use list_del_init so people can check to see if they
8012          * are still on the list after taking the semaphore
8013          */
8014         list_del_init(&block_group->list);
8015         if (list_empty(&block_group->space_info->block_groups[index]))
8016                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8017         up_write(&block_group->space_info->groups_sem);
8018
8019         if (block_group->cached == BTRFS_CACHE_STARTED)
8020                 wait_block_group_cache_done(block_group);
8021
8022         btrfs_remove_free_space_cache(block_group);
8023
8024         spin_lock(&block_group->space_info->lock);
8025         block_group->space_info->total_bytes -= block_group->key.offset;
8026         block_group->space_info->bytes_readonly -= block_group->key.offset;
8027         block_group->space_info->disk_total -= block_group->key.offset * factor;
8028         spin_unlock(&block_group->space_info->lock);
8029
8030         memcpy(&key, &block_group->key, sizeof(key));
8031
8032         btrfs_clear_space_info_full(root->fs_info);
8033
8034         btrfs_put_block_group(block_group);
8035         btrfs_put_block_group(block_group);
8036
8037         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8038         if (ret > 0)
8039                 ret = -EIO;
8040         if (ret < 0)
8041                 goto out;
8042
8043         ret = btrfs_del_item(trans, root, path);
8044 out:
8045         btrfs_free_path(path);
8046         return ret;
8047 }
8048
8049 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8050 {
8051         struct btrfs_space_info *space_info;
8052         struct btrfs_super_block *disk_super;
8053         u64 features;
8054         u64 flags;
8055         int mixed = 0;
8056         int ret;
8057
8058         disk_super = fs_info->super_copy;
8059         if (!btrfs_super_root(disk_super))
8060                 return 1;
8061
8062         features = btrfs_super_incompat_flags(disk_super);
8063         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8064                 mixed = 1;
8065
8066         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8067         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8068         if (ret)
8069                 goto out;
8070
8071         if (mixed) {
8072                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8073                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8074         } else {
8075                 flags = BTRFS_BLOCK_GROUP_METADATA;
8076                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8077                 if (ret)
8078                         goto out;
8079
8080                 flags = BTRFS_BLOCK_GROUP_DATA;
8081                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8082         }
8083 out:
8084         return ret;
8085 }
8086
8087 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8088 {
8089         return unpin_extent_range(root, start, end);
8090 }
8091
8092 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8093                                u64 num_bytes, u64 *actual_bytes)
8094 {
8095         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8096 }
8097
8098 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8099 {
8100         struct btrfs_fs_info *fs_info = root->fs_info;
8101         struct btrfs_block_group_cache *cache = NULL;
8102         u64 group_trimmed;
8103         u64 start;
8104         u64 end;
8105         u64 trimmed = 0;
8106         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8107         int ret = 0;
8108
8109         /*
8110          * try to trim all FS space, our block group may start from non-zero.
8111          */
8112         if (range->len == total_bytes)
8113                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8114         else
8115                 cache = btrfs_lookup_block_group(fs_info, range->start);
8116
8117         while (cache) {
8118                 if (cache->key.objectid >= (range->start + range->len)) {
8119                         btrfs_put_block_group(cache);
8120                         break;
8121                 }
8122
8123                 start = max(range->start, cache->key.objectid);
8124                 end = min(range->start + range->len,
8125                                 cache->key.objectid + cache->key.offset);
8126
8127                 if (end - start >= range->minlen) {
8128                         if (!block_group_cache_done(cache)) {
8129                                 ret = cache_block_group(cache, NULL, root, 0);
8130                                 if (!ret)
8131                                         wait_block_group_cache_done(cache);
8132                         }
8133                         ret = btrfs_trim_block_group(cache,
8134                                                      &group_trimmed,
8135                                                      start,
8136                                                      end,
8137                                                      range->minlen);
8138
8139                         trimmed += group_trimmed;
8140                         if (ret) {
8141                                 btrfs_put_block_group(cache);
8142                                 break;
8143                         }
8144                 }
8145
8146                 cache = next_block_group(fs_info->tree_root, cache);
8147         }
8148
8149         range->len = trimmed;
8150         return ret;
8151 }